mirror of
https://github.com/seaweedfs/seaweedfs.git
synced 2026-09-29 11:15:34 +00:00
b9ad62fc16
* fix 11400 * persist failed-recovery quarantine and harden rollback - record the unavailable state in a .unavailable marker, fsync it, and re-arm it on load so a restart cannot serve an unverified pair - quarantine the volume so heartbeats stop advertising it - block MarkVolumeWritable while unavailable, rechecked under noWriteLock - fail every request of a failed batch, not only the succeeded ones - restore the needle map and truncate .dat on inline fsync rollback failure - add truncateIndex for the sorted-file needle map - mirror the fail-closed semantics in the Rust volume server * volume: erase rolled-back mappings instead of leaving tombstones A rolled-back batch or failed inline write used Delete() to undo a needle that did not exist beforehand, leaving a tombstoned map entry whose stale offset makes the next write to that needle fail reading a header that no longer exists. Add removeMapping/restoreMapping to the mappers so recovery erases entries that were absent before the batch and reinstates the exact prior offset/size for ones that were, including tombstones. The index row still goes through Delete so a replay forgets the needle. * volume: gate bulk readers on unavailable and fsync the marker's dir - fsync_dir(&self.dir) synced the volume dir's parent, not the dir holding .unavailable; pass the marker path so the create survives a host crash - export UnavailableError and check it in ReadAllNeedles, VolumeTailSender, VolumeIncrementalCopy, and IncrementalBackup so replica-sync paths cannot stream or append data from an unverified .dat/.idx pair; mirror on the Rust side via read_dat_slice, read_all_needles, dat_scan_plan, and the incremental-copy handler * volume: drop issue references from comments near touched code * volume: stop active scans when the volume becomes unavailable The stream entry-point checks ran once per RPC, so a volume quarantined by a failed recovery mid-scan kept serving data. Recheck availability per needle/chunk on the detached read paths: tail scan and heartbeat, read-all, incremental copy, incremental backup writes, and the Rust StreamingBody chunk reads. Rust incremental copy also rejects a quarantined volume before sync_to_disk touches the backend. --------- Co-authored-by: Chris Lu <chris.lu@gmail.com>
51 lines
1.3 KiB
Go
51 lines
1.3 KiB
Go
package storage
|
|
|
|
import (
|
|
"github.com/seaweedfs/seaweedfs/weed/pb/volume_server_pb"
|
|
"github.com/seaweedfs/seaweedfs/weed/storage/needle"
|
|
"github.com/seaweedfs/seaweedfs/weed/storage/super_block"
|
|
)
|
|
|
|
type VolumeFileScanner4ReadAll struct {
|
|
Stream volume_server_pb.VolumeServer_ReadAllNeedlesServer
|
|
V *Volume
|
|
}
|
|
|
|
func (scanner *VolumeFileScanner4ReadAll) VisitSuperBlock(superBlock super_block.SuperBlock) error {
|
|
return nil
|
|
|
|
}
|
|
func (scanner *VolumeFileScanner4ReadAll) ReadNeedleBody() bool {
|
|
return true
|
|
}
|
|
|
|
func (scanner *VolumeFileScanner4ReadAll) VisitNeedle(n *needle.Needle, offset int64, needleHeader, needleBody []byte) error {
|
|
|
|
if err := scanner.V.UnavailableError(); err != nil {
|
|
return err
|
|
}
|
|
nv, ok := scanner.V.nm.Get(n.Id)
|
|
if !ok {
|
|
return nil
|
|
}
|
|
if nv.Offset.ToActualOffset() != offset {
|
|
return nil
|
|
}
|
|
|
|
sendErr := scanner.Stream.Send(&volume_server_pb.ReadAllNeedlesResponse{
|
|
VolumeId: uint32(scanner.V.Id),
|
|
NeedleId: uint64(n.Id),
|
|
Cookie: uint32(n.Cookie),
|
|
NeedleBlob: n.Data,
|
|
NeedleBlobCompressed: n.IsCompressed(),
|
|
LastModified: n.LastModified,
|
|
Crc: n.Checksum.Value(),
|
|
Name: n.Name,
|
|
Mime: n.Mime,
|
|
})
|
|
if sendErr != nil {
|
|
return sendErr
|
|
}
|
|
return nil
|
|
}
|