Compare commits

..

4 Commits

Author SHA1 Message Date
马登山 bcc2771dcb fix(filemeta): fence unsafe heal key components 2026-08-22 19:42:47 +08:00
马登山 f40f8179a4 fix(scanner): bound orphan heal discovery fallback 2026-08-22 19:35:56 +08:00
马登山 27a921372c fix(scanner): preserve unversioned heal retries 2026-08-22 19:12:27 +08:00
Zhengchao An 1a3be70d98 fix(ecstore): preserve remote delete error types (#6371) 2026-08-22 17:07:02 +08:00
5 changed files with 878 additions and 80 deletions
+595 -1
View File
@@ -21,6 +21,7 @@ use arc_swap::ArcSwapOption;
use rmp::Marker;
use serde::{Deserialize, Serialize};
use std::cmp::Ordering;
use std::collections::{HashMap, HashSet};
use std::str::from_utf8;
use std::{
fmt::Debug,
@@ -37,8 +38,13 @@ use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt};
use tokio::spawn;
use tokio::sync::Mutex;
use tracing::{debug, warn};
use uuid::Uuid;
const SLASH_SEPARATOR: &str = "/";
pub const MAX_META_CACHE_HEAL_CANDIDATES: usize = 1024;
/// Keep truncation continuations bounded while still giving the scanner a
/// safe object-level retry for versions that did not fit in the candidate set.
pub const MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS: usize = 64;
#[derive(Clone, Debug, Default)]
pub struct MetadataResolutionParams {
@@ -66,6 +72,47 @@ pub struct MetaCacheEntry {
pub reusable: bool,
}
#[derive(Clone, Debug, PartialEq, Eq, Hash)]
pub enum MetaCacheHealCandidateKind {
Object,
DeleteMarker,
UnversionedObject,
}
#[derive(Clone, Debug, PartialEq, Eq, Hash)]
pub struct MetaCacheHealCandidate {
pub object: String,
pub version_id: Option<Uuid>,
pub kind: MetaCacheHealCandidateKind,
/// Number of raw disk entries that carried this validated version.
pub replica_count: usize,
}
impl MetaCacheHealCandidate {
pub fn validated_version(&self) -> Option<Uuid> {
match self.kind {
MetaCacheHealCandidateKind::Object | MetaCacheHealCandidateKind::DeleteMarker => self.version_id,
MetaCacheHealCandidateKind::UnversionedObject => None,
}
}
pub fn is_unversioned(&self) -> bool {
self.kind == MetaCacheHealCandidateKind::UnversionedObject
}
}
#[derive(Clone, Debug, Default, PartialEq, Eq)]
pub struct MetaCacheHealDiscovery {
pub candidates: Vec<MetaCacheHealCandidate>,
pub unverified_count: usize,
pub truncated: bool,
/// Object names whose validated version set exceeded the candidate cap.
/// The scanner retries these names without a version and with destructive
/// healing disabled; this is an explicit bounded continuation, not a
/// version claim.
pub truncated_objects: Vec<String>,
}
impl MetaCacheEntry {
pub fn marshal_msg(&self) -> Result<Vec<u8>> {
let mut wr = Vec::new();
@@ -370,6 +417,184 @@ impl MetaCacheEntries {
})
}
/// Discover validated object/delete-marker versions and safe unversioned
/// inspection candidates in the raw entries without applying read quorum.
/// This is intentionally separate from [`Self::resolve`]: a sub-quorum
/// version is a valid heal target even though it must not participate in
/// normal reads or writes.
///
/// The validated list is bounded and deduplicated by object, version id,
/// and metadata kind; each candidate retains the number of raw disk
/// entries that carried it so callers can classify sub-quorum versions.
/// Entries whose xl.meta cannot be decoded are counted separately for
/// discovery accounting; they never become versionless destructive heal
/// requests and do not consume the validated quota. An
/// [`MetaCacheHealCandidateKind::UnversionedObject`] is always consumed by
/// a non-destructive scanner request.
pub fn discover_heal_candidates(&self, bucket: &str, max_candidates: usize) -> MetaCacheHealDiscovery {
let limit = max_candidates.min(MAX_META_CACHE_HEAL_CANDIDATES);
if limit == 0 || bucket.is_empty() {
return MetaCacheHealDiscovery::default();
}
let mut discovery = MetaCacheHealDiscovery {
candidates: Vec::<MetaCacheHealCandidate>::with_capacity(limit.min(self.0.len())),
unverified_count: 0,
truncated: false,
truncated_objects: Vec::with_capacity(MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS.min(limit)),
};
let mut seen: HashMap<(String, Option<Uuid>, MetaCacheHealCandidateKind), usize> =
HashMap::with_capacity(limit.min(self.0.len()));
for entry in self.0.iter().flatten() {
if !valid_heal_candidate_name(bucket, entry) {
continue;
}
let meta = match FileMeta::load(&entry.metadata) {
Ok(meta) => meta,
Err(_) => {
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
continue;
}
};
let mut entry_seen = HashSet::new();
for shallow in meta.versions {
let version = match shallow.parse_version_meta() {
Ok(version) if version.valid() => version,
Ok(_) | Err(_) => {
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
continue;
}
};
if version.free_version() {
continue;
}
let payload_header = version.header();
if normalize_version_id(shallow.header.version_id) != normalize_version_id(payload_header.version_id)
|| shallow.header.version_type != payload_header.version_type
{
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
continue;
}
let (kind, version_id) = match version.version_type {
VersionType::Object
if version.object.is_some() && version.delete_marker.is_none() && version.legacy_object.is_none() =>
{
match version.object.as_ref().and_then(|object| object.version_id) {
Some(id) if !id.is_nil() => (MetaCacheHealCandidateKind::Object, Some(id)),
Some(_) | None => (MetaCacheHealCandidateKind::UnversionedObject, None),
}
}
VersionType::Delete
if version.delete_marker.is_some() && version.object.is_none() && version.legacy_object.is_none() =>
{
let Some(id) = version.delete_marker.as_ref().and_then(|marker| marker.version_id) else {
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
continue;
};
if id.is_nil() {
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
continue;
}
(MetaCacheHealCandidateKind::DeleteMarker, Some(id))
}
VersionType::Legacy
if version.legacy_object.is_some() && version.object.is_none() && version.delete_marker.is_none() =>
{
let Some(legacy) = version.legacy_object.as_ref() else {
continue;
};
if legacy.version_id.is_empty() {
(MetaCacheHealCandidateKind::UnversionedObject, None)
} else {
let Ok(id) = Uuid::parse_str(&legacy.version_id) else {
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
continue;
};
if id.is_nil() {
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
continue;
}
(MetaCacheHealCandidateKind::Object, Some(id))
}
}
_ => {
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
continue;
}
};
if normalize_version_id(payload_header.version_id) != version_id {
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
continue;
}
// `all_parts=true` is the trust-boundary check for versioned
// candidates. A null/legacy object may still need the old
// non-destructive inspection fallback when its part arrays
// are parseable but incomplete; never use that fallback for
// a candidate carrying a real version id.
let file_info = match version.clone().into_fileinfo(bucket, &entry.name, true) {
Ok(file_info) => file_info,
Err(_) if version_id.is_none() && matches!(kind, MetaCacheHealCandidateKind::UnversionedObject) => {
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
match version.into_fileinfo(bucket, &entry.name, false) {
Ok(file_info) => file_info,
Err(_) => continue,
}
}
Err(_) => {
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
continue;
}
};
if file_info.volume != bucket || file_info.name != entry.name {
discovery.unverified_count = discovery.unverified_count.saturating_add(1);
continue;
}
let candidate = MetaCacheHealCandidate {
object: entry.name.clone(),
version_id,
kind,
replica_count: 1,
};
let key = (candidate.object.clone(), candidate.version_id, candidate.kind.clone());
if entry_seen.contains(&key) {
continue;
}
if let Some(index) = seen.get(&key).copied() {
entry_seen.insert(key);
discovery.candidates[index].replica_count = discovery.candidates[index].replica_count.saturating_add(1);
} else if discovery.candidates.len() >= limit {
// Keep the validated candidate list bounded, but retain a
// bounded object-level continuation so the scanner cannot
// silently lose every version of a busy object.
discovery.truncated = true;
if discovery.truncated_objects.len() < MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS
&& !discovery.truncated_objects.iter().any(|object| object == &candidate.object)
{
discovery.truncated_objects.push(candidate.object.clone());
}
// The remaining versions in this raw entry cannot add a
// bounded candidate; avoid parsing a very long history
// after the safe continuation has been recorded.
break;
} else {
entry_seen.insert(key.clone());
seen.insert(key, discovery.candidates.len());
discovery.candidates.push(candidate);
}
}
}
discovery
}
fn resolve_inner(&self, mut params: MetadataResolutionParams, enforce_write_quorum: bool) -> Option<MetaCacheEntry> {
if self.0.is_empty() {
debug!(
@@ -546,6 +771,33 @@ impl MetaCacheEntries {
}
}
fn valid_heal_candidate_name(bucket: &str, entry: &MetaCacheEntry) -> bool {
if bucket.is_empty()
|| entry.name.is_empty()
|| entry.is_dir()
|| entry.name.contains('\\')
|| entry.name.chars().any(char::is_control)
{
return false;
}
// Validate raw key components without normalizing them. The scanner maps
// accepted keys to filesystem paths later, so dot components and empty
// internal components must be rejected before that boundary. A final
// empty component is retained for valid keys ending in '/'.
let mut components = entry.name.split('/').peekable();
while let Some(component) = components.next() {
if component == "." || component == ".." || (component.is_empty() && components.peek().is_some()) {
return false;
}
}
true
}
fn normalize_version_id(version_id: Option<Uuid>) -> Option<Uuid> {
version_id.filter(|id| !id.is_nil())
}
#[derive(Debug, Default)]
pub struct MetaCacheEntriesSortedResult {
pub entries: Option<MetaCacheEntriesSorted>,
@@ -991,7 +1243,7 @@ impl<T: Clone + Debug + Send + Sync + 'static> Cache<T> {
mod tests {
use super::*;
use crate::test_data::create_real_xlmeta;
use crate::{FileMetaVersion, MetaDeleteMarker, TRANSITION_COMPLETE};
use crate::{FileMetaVersion, MetaDeleteMarker, MetaObjectV1, MetaObjectV1Erasure, MetaObjectV1Stat, TRANSITION_COMPLETE};
use std::collections::HashMap;
use std::io::Cursor;
use std::sync::{
@@ -1592,6 +1844,348 @@ mod tests {
);
}
#[test]
fn discover_heal_candidates_keeps_sub_quorum_versions_and_deduplicates() {
let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
let entries = MetaCacheEntries(vec![
Some(metacache_entry_single_version(1, now, "one")),
Some(metacache_entry_single_version(2, now, "two")),
Some(metacache_entry_single_version(2, now, "two")),
Some(metacache_entry_single_version(3, now, "three")),
]);
let discovery = entries.discover_heal_candidates("bucket", 16);
let ids: std::collections::HashSet<Uuid> = discovery
.candidates
.iter()
.filter_map(|candidate| candidate.version_id)
.collect();
assert_eq!(
ids,
[Uuid::from_u128(1), Uuid::from_u128(2), Uuid::from_u128(3)]
.into_iter()
.collect()
);
assert_eq!(discovery.candidates.len(), 3, "duplicate tied versions must be emitted once");
assert_eq!(
discovery
.candidates
.iter()
.find(|candidate| candidate.version_id == Some(Uuid::from_u128(2)))
.expect("duplicate version should be discovered")
.replica_count,
2
);
}
#[test]
fn discover_heal_candidates_does_not_count_duplicate_versions_within_one_entry() {
let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
let mut meta = FileMeta::load(&metacache_entry_single_version(1, now, "duplicate").metadata)
.expect("duplicate fixture should decode");
meta.versions.push(meta.versions[0].clone());
let entry = MetaCacheEntry {
name: "object".to_string(),
metadata: meta.marshal_msg().expect("duplicate metadata should marshal"),
cached: Some(meta),
reusable: false,
};
let discovery = MetaCacheEntries(vec![Some(entry)]).discover_heal_candidates("bucket", 16);
let candidate = discovery
.candidates
.iter()
.find(|candidate| candidate.version_id == Some(Uuid::from_u128(1)))
.expect("duplicate fixture should be discovered");
assert_eq!(candidate.replica_count, 1);
}
#[test]
fn discover_heal_candidates_covers_divergent_quorum_boundaries_n2_n4_n6() {
let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
for (disk_count, quorum) in [(2usize, 1usize), (4, 2), (6, 3)] {
let target_id = Uuid::from_u128(0x1000 + disk_count as u128);
for target_replicas in [quorum.saturating_sub(1), quorum, quorum + 1] {
let entries = (0..disk_count)
.map(|disk| {
let version_id = if disk < target_replicas {
target_id
} else {
Uuid::from_u128(0x2000 + disk as u128)
};
Some(metacache_entry_single_version(version_id.as_u128(), now, "divergent"))
})
.collect();
let discovery = MetaCacheEntries(entries).discover_heal_candidates("bucket", 32);
let target = discovery
.candidates
.iter()
.find(|candidate| candidate.version_id == Some(target_id));
assert_eq!(target.is_some(), target_replicas > 0, "N={disk_count}, replicas={target_replicas}");
if let Some(target) = target {
assert_eq!(target.replica_count, target_replicas);
}
}
}
}
#[test]
fn discover_heal_candidates_separates_delete_markers_and_preserves_unversioned_objects() {
let mut marker_meta = FileMeta::new();
marker_meta
.add_version(FileInfo {
volume: "bucket".to_string(),
name: "object".to_string(),
version_id: Some(Uuid::from_u128(99)),
deleted: true,
mod_time: Some(OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp")),
..Default::default()
})
.expect("delete marker should be added");
let marker = MetaCacheEntry {
name: "object".to_string(),
metadata: marker_meta.marshal_msg().expect("delete marker metadata should marshal"),
cached: Some(marker_meta),
reusable: false,
};
let unversioned_entry = metacache_entry_with_mod_time(
OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp"),
"unversioned",
);
let discovery = MetaCacheEntries(vec![Some(marker), Some(unversioned_entry)]).discover_heal_candidates("bucket", 16);
assert!(discovery.candidates.iter().any(|candidate| {
candidate.kind == MetaCacheHealCandidateKind::DeleteMarker && candidate.version_id == Some(Uuid::from_u128(99))
}));
assert!(discovery.candidates.iter().any(|candidate| {
candidate.kind == MetaCacheHealCandidateKind::UnversionedObject && candidate.version_id.is_none()
}));
}
#[test]
fn discover_heal_candidates_rejects_delete_markers_without_ids() {
let mut marker_meta = FileMeta::new();
marker_meta
.add_version(FileInfo {
volume: "bucket".to_string(),
name: "object".to_string(),
deleted: true,
mod_time: Some(OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp")),
..Default::default()
})
.expect("nil delete marker should be added");
let discovery = MetaCacheEntries(vec![Some(MetaCacheEntry {
name: "object".to_string(),
metadata: marker_meta.marshal_msg().expect("nil marker metadata should marshal"),
cached: Some(marker_meta),
reusable: false,
})])
.discover_heal_candidates("bucket", 16);
assert!(
!discovery
.candidates
.iter()
.any(|candidate| candidate.kind == MetaCacheHealCandidateKind::DeleteMarker)
);
assert!(discovery.unverified_count >= 1);
}
#[test]
fn discover_heal_candidates_skips_free_versions() {
let object_id = Uuid::from_u128(100);
let free_id = Uuid::from_u128(101);
let mut meta = FileMeta::new();
meta.add_version(FileInfo {
volume: "bucket".to_string(),
name: "object".to_string(),
version_id: Some(object_id),
transition_status: TRANSITION_COMPLETE.to_string(),
transitioned_objname: "remote/object".to_string(),
transition_version_id: Some(Uuid::from_u128(102)),
transition_tier: "WARM".to_string(),
mod_time: Some(OffsetDateTime::now_utc()),
..Default::default()
})
.expect("transitioned object should be added");
let mut delete = FileInfo {
volume: "bucket".to_string(),
name: "object".to_string(),
version_id: Some(object_id),
mod_time: Some(OffsetDateTime::now_utc()),
..Default::default()
};
delete.set_tier_free_version_id(&free_id.to_string());
meta.delete_version(&delete).expect("free version should be persisted");
let discovery = MetaCacheEntries(vec![Some(MetaCacheEntry {
name: "object".to_string(),
metadata: meta.marshal_msg().expect("free version metadata should marshal"),
cached: Some(meta),
reusable: false,
})])
.discover_heal_candidates("bucket", 16);
assert!(discovery.candidates.is_empty());
}
#[test]
fn discover_heal_candidates_preserves_unversioned_legacy_object() {
let legacy = MetaObjectV1 {
version: "1.0.1".to_string(),
format: "xl".to_string(),
stat: MetaObjectV1Stat {
size: 1,
mod_time: Some(OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp")),
name: "object".to_string(),
..Default::default()
},
erasure: MetaObjectV1Erasure {
data_blocks: 4,
parity_blocks: 2,
index: 1,
distribution: vec![1, 2, 3, 4, 5, 6],
..Default::default()
},
..Default::default()
};
let version = FileMetaVersion {
version_type: VersionType::Legacy,
legacy_object: Some(legacy),
..Default::default()
};
let mut meta = FileMeta::new();
meta.versions
.push(FileMetaShallowVersion::try_from(version).expect("legacy metadata should marshal"));
let discovery = MetaCacheEntries(vec![Some(MetaCacheEntry {
name: "object".to_string(),
metadata: meta.marshal_msg().expect("legacy metadata should marshal"),
cached: Some(meta),
reusable: false,
})])
.discover_heal_candidates("bucket", 16);
assert!(discovery.candidates.iter().any(|candidate| {
candidate.kind == MetaCacheHealCandidateKind::UnversionedObject && candidate.version_id.is_none()
}));
}
#[test]
fn discover_heal_candidates_rejects_nil_and_malformed_metadata_and_is_bounded() {
let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
let mut nil = metacache_entry_single_version(1, now, "nil");
let mut nil_meta = FileMeta::load(&nil.metadata).expect("nil fixture should decode");
let mut nil_version = nil_meta.versions[0]
.parse_version_meta()
.expect("nil fixture version should decode");
nil_version.object.as_mut().expect("object fixture").version_id = Some(Uuid::nil());
nil_meta.versions[0] = FileMetaShallowVersion::try_from(nil_version).expect("nil fixture should marshal");
nil.metadata = nil_meta.marshal_msg().expect("nil fixture metadata should marshal");
let mut mismatched = metacache_entry_single_version(2, now, "mismatched");
let mut mismatched_meta = FileMeta::load(&mismatched.metadata).expect("mismatched fixture should decode");
mismatched_meta.versions[0].header.version_id = Some(Uuid::from_u128(200));
mismatched.metadata = mismatched_meta.marshal_msg().expect("mismatched metadata should marshal");
let mut short_parts = metacache_entry_single_version(3, now, "short-parts");
let mut short_parts_meta = FileMeta::load(&short_parts.metadata).expect("short-parts fixture should decode");
let mut short_parts_version = short_parts_meta.versions[0]
.parse_version_meta()
.expect("short-parts fixture version should decode");
let object = short_parts_version.object.as_mut().expect("object fixture");
object.part_numbers = vec![1];
object.part_actual_sizes = vec![1];
object.part_sizes.clear();
short_parts_meta.versions[0] = FileMetaShallowVersion::try_from(short_parts_version).expect("short-parts should marshal");
short_parts.metadata = short_parts_meta.marshal_msg().expect("short-parts metadata should marshal");
let mut short_unversioned = metacache_entry_with_mod_time(now, "short-unversioned");
let mut short_unversioned_meta =
FileMeta::load(&short_unversioned.metadata).expect("short-unversioned fixture should decode");
let mut short_unversioned_version = short_unversioned_meta.versions[0]
.parse_version_meta()
.expect("short-unversioned version should decode");
let unversioned_object = short_unversioned_version.object.as_mut().expect("unversioned object fixture");
unversioned_object.part_numbers = vec![1];
unversioned_object.part_actual_sizes = vec![1];
unversioned_object.part_sizes.clear();
short_unversioned_meta.versions[0] =
FileMetaShallowVersion::try_from(short_unversioned_version).expect("short-unversioned should marshal");
short_unversioned.metadata = short_unversioned_meta
.marshal_msg()
.expect("short-unversioned metadata should marshal");
let mut malformed = nil.clone();
malformed.name = "malformed".to_string();
malformed.metadata = vec![1, 2, 3];
let entries = MetaCacheEntries(
std::iter::once(Some(nil))
.chain(std::iter::once(Some(mismatched)))
.chain(std::iter::once(Some(short_parts)))
.chain(std::iter::once(Some(short_unversioned)))
.chain(std::iter::once(Some(malformed)))
.chain((0..32).map(|id| Some(metacache_entry_single_version(id + 10, now, "bounded"))))
.collect(),
);
let discovery = entries.discover_heal_candidates("bucket", 5);
assert!(discovery.candidates.len() <= 5);
assert!(discovery.truncated, "bounded discovery must expose dropped candidates");
assert!(
discovery.truncated_objects.iter().any(|object| object == "object"),
"bounded discovery must expose an object-level safe continuation"
);
assert!(
!discovery
.candidates
.iter()
.any(|candidate| candidate.version_id == Some(Uuid::nil()))
);
assert!(
!discovery
.candidates
.iter()
.any(|candidate| candidate.version_id == Some(Uuid::from_u128(2)))
);
assert!(
!discovery
.candidates
.iter()
.any(|candidate| candidate.version_id == Some(Uuid::from_u128(3)))
);
assert!(discovery.candidates.iter().any(|candidate| {
candidate.kind == MetaCacheHealCandidateKind::UnversionedObject && candidate.version_id.is_none()
}));
assert!(
discovery.unverified_count >= 1,
"malformed and rejected metadata must remain observable during discovery"
);
for invalid_name in [
"../object",
"./object",
"object/../other",
"object//name",
"object\\name",
"object\u{0001}name",
"object\0name",
] {
let mut entry = metacache_entry_single_version(400, now, invalid_name);
entry.name = invalid_name.to_string();
let discovery = MetaCacheEntries(vec![Some(entry)]).discover_heal_candidates("bucket", 5);
assert!(
discovery.candidates.is_empty(),
"invalid key should not become a heal candidate: {invalid_name:?}"
);
}
for valid_name in ["trailing/", "prefix/object"] {
let mut entry = metacache_entry_single_version(401, now, valid_name);
entry.name = valid_name.to_string();
let discovery = MetaCacheEntries(vec![Some(entry)]).discover_heal_candidates("bucket", 5);
assert_eq!(discovery.candidates.len(), 1, "raw S3 key should remain opaque: {valid_name:?}");
}
}
#[test]
fn resolve_rejects_partial_latest_and_returns_committed_previous_metadata() {
let old_mod_time = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
+120 -66
View File
@@ -43,7 +43,10 @@ use rustfs_common::metrics::{
UpdateCurrentPathFn, current_path_updater, global_metrics,
};
use rustfs_common::trace_bus::{TraceEvent, TraceFunc, TraceKind, trace_emit, trace_subscriber_count};
use rustfs_filemeta::{MetaCacheEntries, MetaCacheEntry, MetadataResolutionParams};
use rustfs_filemeta::{
MAX_META_CACHE_HEAL_CANDIDATES, MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS, MetaCacheEntries, MetaCacheEntry,
MetaCacheHealCandidateKind,
};
use rustfs_utils::path::{SLASH_SEPARATOR, path_join_buf};
use s3s::dto::{BucketLifecycleConfiguration, ObjectLockConfiguration, VersioningConfiguration};
use time::OffsetDateTime;
@@ -96,6 +99,11 @@ const METRIC_SCANNER_EXCESS_OBJECT_VERSION_SIZE_TOTAL: &str = "rustfs_scanner_ex
const METRIC_SCANNER_EXCESS_FOLDERS_TOTAL: &str = "rustfs_scanner_excess_folders_total";
const METRIC_SCANNER_PENDING_HEAL_PRUNE_TOTAL: &str = "rustfs_scanner_pending_heal_prune_total";
const METRIC_SCANNER_PENDING_HEAL_MALFORMED_TOTAL: &str = "rustfs_scanner_pending_heal_malformed_total";
const METRIC_SCANNER_HEAL_DISCOVERY_CANDIDATES_TOTAL: &str = "rustfs_scanner_heal_discovery_candidates_total";
const METRIC_SCANNER_HEAL_DISCOVERY_SUB_QUORUM_TOTAL: &str = "rustfs_scanner_heal_discovery_sub_quorum_total";
const METRIC_SCANNER_HEAL_DISCOVERY_UNVERIFIED_TOTAL: &str = "rustfs_scanner_heal_discovery_unverified_total";
const METRIC_SCANNER_HEAL_DISCOVERY_QUEUED_TOTAL: &str = "rustfs_scanner_heal_discovery_queued_total";
const METRIC_SCANNER_HEAL_DISCOVERY_TRUNCATED_TOTAL: &str = "rustfs_scanner_heal_discovery_truncated_total";
const MAX_PENDING_SCANNER_HEAL_RETRIES_PER_BUCKET: usize = 128;
// --- scanner excess alerts as S3 notification events (rustfs/backlog#1868) --
@@ -883,7 +891,7 @@ impl FolderScanner {
object: Option<String>,
version_id: Option<String>,
request: HealChannelRequest,
) -> Result<(), ScannerError> {
) -> Result<HealAdmissionResult, ScannerError> {
let candidate_type = pending_scanner_heal_candidate_type(kind);
let priority = request.priority;
let scan_mode = request.scan_mode.unwrap_or(self.scan_mode);
@@ -911,7 +919,7 @@ impl FolderScanner {
error = %err,
"Scanner deferred heal request after channel error"
);
return Ok(());
return Ok(HealAdmissionResult::Full);
}
};
self.update_pending_scanner_heal_after_admission(
@@ -923,7 +931,7 @@ impl FolderScanner {
result,
);
if result.is_admitted() {
return Ok(());
return Ok(result);
}
record_high_priority_heal_escalation(candidate_type, priority, result);
@@ -944,7 +952,7 @@ impl FolderScanner {
state = "high_priority_not_admitted",
"Scanner high-priority heal admission failed"
);
Ok(())
Ok(result)
}
pub fn set_heal_object_select(&mut self, prob: u32) {
@@ -1736,14 +1744,7 @@ impl FolderScanner {
break;
}
let mut resolver = MetadataResolutionParams {
dir_quorum: self.disks_quorum,
obj_quorum: self.disks_quorum,
bucket: "".to_string(),
strict: false,
..Default::default()
};
let mut previous_bucket = String::new();
for name in abandoned_children {
if !self.should_heal().await {
break;
@@ -1751,7 +1752,7 @@ impl FolderScanner {
let (bucket, prefix) = path2_bucket_object(name.as_str());
if bucket != resolver.bucket {
if bucket != previous_bucket {
self.send_required_scanner_heal_request(
PendingScannerHealKind::Bucket,
bucket.clone(),
@@ -1760,10 +1761,9 @@ impl FolderScanner {
build_bucket_heal_request(bucket.clone(), HealChannelPriority::High),
)
.await?;
previous_bucket = bucket.clone();
}
resolver.bucket = bucket.clone();
let child_ctx = ctx.child_token();
let (agreed_tx, mut agreed_rx) = mpsc::channel::<String>(1);
@@ -1880,6 +1880,8 @@ impl FolderScanner {
let mut agreed_closed = false;
let mut partial_closed = false;
let mut finished_closed = false;
let mut seen_heal_candidates: HashSet<(String, Option<String>, MetaCacheHealCandidateKind)> = HashSet::new();
let mut seen_truncated_objects: HashSet<String> = HashSet::new();
loop {
if agreed_closed && partial_closed && finished_closed {
@@ -1904,65 +1906,117 @@ impl FolderScanner {
break;
}
let Some(entry) = resolve_object_heal_entry(&entries, resolver.clone()) else {
continue;
};
(self.update_current_path)(&entry.name).await;
if entry.is_dir() {
continue;
let discovery = entries.discover_heal_candidates(&bucket, MAX_META_CACHE_HEAL_CANDIDATES);
counter!(METRIC_SCANNER_HEAL_DISCOVERY_CANDIDATES_TOTAL)
.increment(u64::try_from(discovery.candidates.len()).unwrap_or(u64::MAX));
counter!(METRIC_SCANNER_HEAL_DISCOVERY_SUB_QUORUM_TOTAL).increment(
u64::try_from(
discovery
.candidates
.iter()
.filter(|candidate| candidate.replica_count < disks_quorum)
.count(),
)
.unwrap_or(u64::MAX),
);
counter!(METRIC_SCANNER_HEAL_DISCOVERY_UNVERIFIED_TOTAL).increment(
u64::try_from(discovery.unverified_count).unwrap_or(u64::MAX),
);
if discovery.truncated {
counter!(METRIC_SCANNER_HEAL_DISCOVERY_TRUNCATED_TOTAL).increment(1);
}
let fivs = match entry.file_info_versions(&bucket) {
Ok(fivs) => fivs,
Err(e) => {
error!(
target: "rustfs::scanner::folder",
event = EVENT_SCANNER_FOLDER_STATE,
component = LOG_COMPONENT_SCANNER,
subsystem = LOG_SUBSYSTEM_FOLDER,
bucket = %bucket,
entry = %entry.name,
state = "file_info_versions_failed",
error = %e,
"Scanner list_path_raw failed to resolve file versions"
);
self.send_required_scanner_heal_request(
PendingScannerHealKind::Object,
bucket.clone(),
Some(entry.name.clone()),
None,
build_object_heal_request(
bucket.clone(),
entry.name.clone(),
None,
self.scan_mode,
HealChannelPriority::High,
),
)
.await?;
found_objects = true;
for candidate in discovery.candidates {
let sub_quorum_candidate = candidate.replica_count < disks_quorum;
let version_id = candidate.validated_version().map(|id| id.to_string());
let identity = (candidate.object.clone(), version_id.clone(), candidate.kind.clone());
if seen_heal_candidates.len() >= MAX_META_CACHE_HEAL_CANDIDATES
&& !seen_heal_candidates.contains(&identity)
{
continue;
}
};
for fiv in fivs.versions {
let version_id = fiv.version_id.and_then(|v| if v.is_nil() { None } else { Some(v.to_string()) });
self.send_required_scanner_heal_request(
PendingScannerHealKind::Object,
bucket.clone(),
Some(entry.name.clone()),
version_id.clone(),
build_object_heal_request(
if !seen_heal_candidates.insert(identity) {
continue;
}
let request = if candidate.is_unversioned() {
build_non_destructive_object_heal_request(
bucket.clone(),
entry.name.clone(),
version_id,
candidate.object.clone(),
self.scan_mode,
HealChannelPriority::High,
),
)
} else {
build_object_heal_request(
bucket.clone(),
candidate.object.clone(),
version_id.clone(),
self.scan_mode,
HealChannelPriority::High,
)
};
(self.update_current_path)(&candidate.object).await;
let admission = self.send_required_scanner_heal_request(
PendingScannerHealKind::Object,
bucket.clone(),
Some(candidate.object.clone()),
version_id.clone(),
request,
)
.await?;
if admission.is_admitted() {
counter!(METRIC_SCANNER_HEAL_DISCOVERY_QUEUED_TOTAL).increment(1);
} else if sub_quorum_candidate {
self.mark_pending_scanner_heal_reason(
PendingScannerHealKind::Object,
&bucket,
Some(&candidate.object),
version_id.as_deref(),
"sub_quorum_metadata",
);
}
found_objects = true;
}
// A bounded candidate union may overflow for an
// object with a very long version history. Keep
// that overflow explicit and issue one safe,
// versionless inspection request per object so
// the dropped versions are not silently treated
// as absent. This continuation is deliberately
// outside the versioned candidate cap and always
// disables destructive cleanup.
for object in discovery.truncated_objects {
if seen_truncated_objects.len() >= MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS
&& !seen_truncated_objects.contains(&object)
{
continue;
}
if !seen_truncated_objects.insert(object.clone()) {
continue;
}
let identity = (object.clone(), None, MetaCacheHealCandidateKind::UnversionedObject);
if !seen_heal_candidates.insert(identity) {
continue;
}
let request = build_non_destructive_object_heal_request(
bucket.clone(),
object.clone(),
self.scan_mode,
HealChannelPriority::High,
);
(self.update_current_path)(&object).await;
let admission = self
.send_required_scanner_heal_request(
PendingScannerHealKind::Object,
bucket.clone(),
Some(object.clone()),
None,
request,
)
.await?;
if admission.is_admitted() {
counter!(METRIC_SCANNER_HEAL_DISCOVERY_QUEUED_TOTAL).increment(1);
}
found_objects = true;
}
@@ -13,6 +13,8 @@
// limitations under the License.
/// Per-object scan actions: ScannerItem, the get-size failure policy, and the heal/ILM admission helpers.
use super::*;
#[cfg(test)]
use rustfs_filemeta::MetadataResolutionParams;
/// Cached folder information for scanning
#[derive(Clone, Debug)]
@@ -88,6 +90,22 @@ pub(super) fn build_object_heal_request(
}
}
/// Build the versionless inspection request used when discovery cannot prove
/// a destructive version identity (for example an unversioned object or a
/// bounded candidate overflow). The explicit flag is the fail-closed safety
/// boundary; callers must not reconstruct it with the destructive default.
pub(super) fn build_non_destructive_object_heal_request(
bucket: String,
object: String,
scan_mode: HealScanMode,
priority: HealChannelPriority,
) -> HealChannelRequest {
let mut request = build_object_heal_request(bucket, object, None, scan_mode, priority);
request.remove_corrupted = Some(false);
request
}
#[cfg(test)]
pub(super) fn resolve_object_heal_entry(
entries: &MetaCacheEntries,
resolver: MetadataResolutionParams,
+39 -7
View File
@@ -105,6 +105,29 @@ impl FolderScanner {
}
}
/// Preserve the discovery reason when a candidate could not be admitted
/// immediately. The existing string field is intentionally reused so the
/// scanner's map-encoded cache schema stays backward compatible.
pub(super) fn mark_pending_scanner_heal_reason(
&mut self,
kind: PendingScannerHealKind,
bucket: &str,
object: Option<&str>,
version_id: Option<&str>,
reason: &str,
) {
if let Some(entry) = self
.new_cache
.info
.pending_heals
.iter_mut()
.find(|entry| pending_scanner_heal_matches(entry, kind, bucket, object, version_id))
{
entry.last_admission_reason = reason.to_string();
self.sync_pending_heals();
}
}
pub(super) fn prune_pending_scanner_heals(&mut self) {
let now = Self::now_secs();
let before_expiry = self.new_cache.info.pending_heals.len();
@@ -305,13 +328,22 @@ pub(super) fn build_pending_scanner_heal_request(entry: &PendingScannerHeal) ->
match entry.kind {
PendingScannerHealKind::Bucket => Some(build_bucket_heal_request(entry.bucket.clone(), HealChannelPriority::High)),
PendingScannerHealKind::Object => entry.object.as_ref().map(|object| {
build_object_heal_request(
entry.bucket.clone(),
object.clone(),
entry.version_id.clone(),
entry.scan_mode,
HealChannelPriority::High,
)
if entry.version_id.is_none() {
build_non_destructive_object_heal_request(
entry.bucket.clone(),
object.clone(),
entry.scan_mode,
HealChannelPriority::High,
)
} else {
build_object_heal_request(
entry.bucket.clone(),
object.clone(),
entry.version_id.clone(),
entry.scan_mode,
HealChannelPriority::High,
)
}
}),
}
}
+106 -6
View File
@@ -17,7 +17,7 @@ use crate::SCANNER_SLEEPER;
use super::*;
use crate::storage_api::VersionPurgeStatusType;
use crate::{DiskOption, Endpoint, STORAGE_FORMAT_FILE, TierStats, new_disk, storageclass};
use rustfs_filemeta::{FileInfo, FileMeta};
use rustfs_filemeta::{FileInfo, FileMeta, MetadataResolutionParams};
use std::io::Write;
#[cfg(unix)]
use std::os::unix::fs::{PermissionsExt, symlink};
@@ -982,6 +982,21 @@ fn test_build_object_heal_request_omits_nil_version_id() {
assert_eq!(request.recreate_missing, Some(false));
}
#[test]
fn test_build_non_destructive_object_heal_request_disables_removal() {
let request = build_non_destructive_object_heal_request(
"bucket".to_string(),
"path/to/object".to_string(),
HealScanMode::Deep,
HealChannelPriority::High,
);
assert_eq!(request.object_version_id, None);
assert_eq!(request.remove_corrupted, Some(false));
assert_eq!(request.recreate_missing, Some(false));
assert_eq!(request.source, HealRequestSource::Scanner);
}
#[test]
fn test_build_bucket_heal_request_disables_recreate_for_scanner() {
let request = build_bucket_heal_request("bucket".to_string(), HealChannelPriority::Low);
@@ -1121,6 +1136,42 @@ fn test_pending_heal_reconstructs_object_request_with_version() {
assert_eq!(request.source, HealRequestSource::Scanner);
}
#[test]
fn test_pending_heal_reconstructs_unversioned_request_without_removal() {
let pending = pending_heal(PendingScannerHealKind::Object, "bucket", Some("object"), None, 1, 1);
let request = build_pending_scanner_heal_request(&pending).expect("unversioned object request should rebuild");
assert!(request.object_version_id.is_none());
assert_eq!(request.remove_corrupted, Some(false));
assert_eq!(request.recreate_missing, Some(false));
}
#[tokio::test]
async fn test_pending_heal_reason_preserves_sub_quorum_discovery() {
let (mut scanner, temp_dir) = build_test_scanner().await;
let _guard = TestGuard::new(u64::MAX, usize::MAX, &mut scanner, temp_dir);
scanner.update_pending_scanner_heal_after_admission(
PendingScannerHealKind::Object,
"bucket",
Some("object"),
Some("version-a"),
HealScanMode::Deep,
HealAdmissionResult::Full,
);
scanner.mark_pending_scanner_heal_reason(
PendingScannerHealKind::Object,
"bucket",
Some("object"),
Some("version-a"),
"sub_quorum_metadata",
);
assert_eq!(scanner.new_cache.info.pending_heals.len(), 1);
assert_eq!(scanner.new_cache.info.pending_heals[0].last_admission_reason, "sub_quorum_metadata");
}
#[test]
fn test_pending_heal_retry_candidates_respect_cap_and_order() {
let pending: Vec<PendingScannerHeal> = (0..(MAX_PENDING_SCANNER_HEAL_RETRIES_PER_BUCKET + 2))
@@ -1323,6 +1374,20 @@ fn metadata_for_object(bucket: &str, object: &str) -> Vec<u8> {
meta.marshal_msg().expect("test metadata should marshal")
}
fn metadata_for_object_version(bucket: &str, object: &str, version_id: Option<Uuid>) -> Vec<u8> {
let mut file_info = FileInfo::new(object, 4, 2);
file_info.volume = bucket.to_string();
file_info.name = object.to_string();
file_info.version_id = version_id;
file_info.versioned = version_id.is_some();
file_info.mod_time = Some(OffsetDateTime::now_utc());
file_info.size = 1;
let mut meta = FileMeta::new();
meta.add_version(file_info).expect("test metadata version should be accepted");
meta.marshal_msg().expect("test metadata should marshal")
}
async fn write_test_object_metadata(root: &std::path::Path, bucket: &str, object: &str) {
write_test_object_metadata_bytes(root, bucket, object, &metadata_for_object(bucket, object)).await;
}
@@ -1726,12 +1791,21 @@ async fn test_scan_folder_exits_when_abandoned_child_listing_finishes() {
let _guard = TestGuard::new(60, 100, &mut scanner, temp_dir.clone());
let heal_starts = Arc::new(AtomicUsize::new(0));
let heal_starts_clone = heal_starts.clone();
let healed_versions = Arc::new(Mutex::new(Vec::<Option<String>>::new()));
let healed_versions_clone = healed_versions.clone();
let mut heal_rx =
rustfs_common::heal_channel::init_heal_channel().expect("heal channel should initialize once for scanner tests");
let _heal_responder = tokio::spawn(async move {
while let Some(command) = heal_rx.recv().await {
if let rustfs_common::heal_channel::HealChannelCommand::Start { response_tx, .. } = command {
if let rustfs_common::heal_channel::HealChannelCommand::Start {
request, response_tx, ..
} = command
{
heal_starts_clone.fetch_add(1, Ordering::Relaxed);
healed_versions_clone
.lock()
.expect("heal version capture lock should not be poisoned")
.push(request.object_version_id);
let _ = response_tx.send(Ok(HealAdmissionResult::Accepted));
}
}
@@ -1739,13 +1813,18 @@ async fn test_scan_folder_exits_when_abandoned_child_listing_finishes() {
let bucket = "src-archive";
let object = "snapshots/37b3f20d941e2f5e6d99114d9bb2f3e67a8a2e5c9c4c5a1b0d6e7f8091a2b3c4";
let metadata = metadata_for_object(bucket, object);
write_test_object_metadata_bytes(&temp_dir, bucket, object, &metadata).await;
let orphan_version = Uuid::from_u128(0x1934);
let shared_version = Uuid::from_u128(0x1935);
let orphan_metadata = metadata_for_object_version(bucket, object, Some(orphan_version));
let shared_metadata = metadata_for_object_version(bucket, object, Some(shared_version));
write_test_object_metadata_bytes(&temp_dir, bucket, object, &orphan_metadata).await;
let mut expected_metadata = vec![(temp_dir.join(bucket).join(object).join("xl.meta"), orphan_metadata.clone())];
let mut disks = vec![scanner.local_disk.clone()];
for disk_name in ["disk2", "disk3", "disk4"] {
let disk_root = temp_dir.join(disk_name);
write_test_object_metadata_bytes(&disk_root, bucket, object, &metadata).await;
write_test_object_metadata_bytes(&disk_root, bucket, object, &shared_metadata).await;
expected_metadata.push((disk_root.join(bucket).join(object).join("xl.meta"), shared_metadata.clone()));
let endpoint = Endpoint::try_from(disk_root.to_string_lossy().as_ref()).expect("failed to create extra disk endpoint");
let disk = new_disk(
&endpoint,
@@ -1794,8 +1873,29 @@ async fn test_scan_folder_exits_when_abandoned_child_listing_finishes() {
.new_cache
.checked_flatten(bucket)
.expect("healed cache must contain canonical child links");
assert_eq!(root.objects, 1);
// The fixture intentionally exposes two divergent version histories, so
// the scanner keeps both logical versions visible while discovering heals.
assert_eq!(root.objects, 2);
assert!(heal_starts.load(Ordering::Relaxed) > 0, "test must execute the heal child-link path");
let orphan_version_text = orphan_version.to_string();
assert!(
healed_versions
.lock()
.expect("heal version capture lock should not be poisoned")
.iter()
.any(|version| version.as_deref() == Some(orphan_version_text.as_str())),
"sub-quorum orphan version must be submitted as an exact heal candidate"
);
for (path, expected) in expected_metadata {
assert_eq!(
tokio::fs::read(&path)
.await
.expect("scanner discovery must not delete metadata"),
expected,
"scanner discovery must not modify candidate metadata: {}",
path.display()
);
}
}
#[tokio::test]