overhaul thread model and backup reliability

Thread model:
  - replace the IOState CAS state machine with a StampedLock: region IO
    (RegionFileStorage.write/flush, now both guarded via @WrapMethod) holds
    the read lock, backups hold the write lock; the old witness-restore
    pattern could wedge the state permanently
  - run saveEverything on the main thread (was called from the backup
    executor), with shutdown-aware waiting; skip autoSave/saveEverything
    while a backup holds the write lock so the main thread never blocks
  - broadcast player messages from the main thread, honor `broadcast`

  Backup correctness:
  - watcher reported the parent directory instead of the changed file, so
    every incremental backup was empty; change sets are now deduplicated,
    atomically drained and merged back on failure; OVERFLOW forces a full
    backup; watch roots are re-registered when their inode changes (world
    format conversion orphaned the watches)
  - fix the region reassembly pipeline: zlib was "decompressed" with
    DeflaterInputStream, writeChunk dropped the 5-byte chunk header,
    the length field was off by one, coordinates were transposed, and
    sector padding was wrong on 4096-byte multiples — every reassembled
    chunk in old bundles is unrecoverable garbage
  - full backups now collect the file list after saving, off-thread

  New features:
  - reflink (FICLONE) staging via FFM: on CoW filesystems the write lock
    is only held while cloning, packing/uploading runs unlocked
  - chained restore: restoring an incremental automatically applies its
    full base and every incremental in between, in order
  - generation-based rotation implementing keepBackups; whole generations
    are deleted so an incremental never loses its base
  - Backup records carry subject/createdAt/incremental for chaining

  Reliability:
  - atomic index persistence (tmp + ATOMIC_MOVE), ConcurrentHashMap for
    the tracked-backup index, fsync bundles before storing
  - BundleReader: overlay extraction for chains, zip-slip guard, fix
    reversed createSymbolicLink args, EOF instead of infinite loop on
    truncated bundles
  - LocalBackupStrategy.deleteBackup compared the wrong field and could
    never delete; S3 retries now catch SdkException (retry was dead code),
    drop ifNoneMatch for idempotent retries
  - ordered shutdown: SERVER_STOPPING awaits in-flight backups before the
    final world save; SERVER_STOPPED runs restore, closes storage and
    cleans the temp dir with isolated error handling
  - fix /backup suspend being registered under /backup create

  Tests: region round-trip byte-comparison, BundleReader overlay/zip-slip/
  symlink/truncation, backup chain resolution and rotation.

  Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
iceBear67 committed 2026-07-17 19:35:43 +08:00
1 parent 592c79ed90
commit 34f7d15566
25 files changed
+1418 -367

No files matched your search

+7
View File
@@ -71,6 +71,13 @@ dependencies {
// https://mvnrepository.com/artifact/software.amazon.awssdk/s3
shadow project(":bundler")
shadow 'software.amazon.awssdk:s3:2.34.0'
testImplementation 'org.junit.jupiter:junit-jupiter:5.13.4'
testRuntimeOnly 'org.junit.platform:junit-platform-launcher'
}
test {
useJUnitPlatform()
}
configurations.implementation.extendsFrom(configurations.shadow)
shadowJar {