seaweedfs/weed/shell/command_volume_fsck.go

package shell

import (
	"bufio"
	"bytes"
	"context"
	"errors"
	"flag"
	"fmt"
	"io"
	"math"
	"net/http"
	"net/url"
	"os"
	"path"
	"path/filepath"
	"sort"
	"strconv"
	"strings"
	"sync"
	"sync/atomic"
	"time"

	"github.com/seaweedfs/seaweedfs/weed/filer"
	"github.com/seaweedfs/seaweedfs/weed/operation"
	"github.com/seaweedfs/seaweedfs/weed/pb"
	"github.com/seaweedfs/seaweedfs/weed/pb/filer_pb"
	"github.com/seaweedfs/seaweedfs/weed/pb/master_pb"
	"github.com/seaweedfs/seaweedfs/weed/pb/volume_server_pb"
	"github.com/seaweedfs/seaweedfs/weed/security"
	"github.com/seaweedfs/seaweedfs/weed/storage"
	"github.com/seaweedfs/seaweedfs/weed/storage/needle"
	"github.com/seaweedfs/seaweedfs/weed/storage/needle_map"
	"github.com/seaweedfs/seaweedfs/weed/storage/types"
	"github.com/seaweedfs/seaweedfs/weed/util"
	util_http "github.com/seaweedfs/seaweedfs/weed/util/http"
	"github.com/seaweedfs/seaweedfs/weed/util/wildcard"
	"golang.org/x/sync/errgroup"
)

func init() {
	Commands = append(Commands, &commandVolumeFsck{})
}

const (
	readbufferSize                 = 16
	jwtFilerTokenExpirationSeconds = 300
	// mtime is only second resolution, so a directory touched this recently is
	// left for the next run rather than compared against a timestamp a
	// concurrent write could share
	directoryQuietPeriod = 5 * time.Second
)

type commandVolumeFsck struct {
	env                       *CommandEnv
	writer                    io.Writer
	bucketsPath               string
	collection                *string
	collectionMatcher         *wildcard.CollectionMatcher
	knownVolumeIds            map[uint32]bool
	volumeIds                 map[uint32]bool
	scopedFilerPath           string
	tempFolder                string
	verbose                   *bool
	forcePurging              *bool
	purgeConcurrency          *int
	skipEcVolumes             *bool
	findMissingChunksInFiler  *bool
	verifyNeedle              *bool
	filerSigningKey           string
	unresolvedManifestEntries atomic.Int64
	purgedDirsLock            sync.Mutex
	purgedDirs                map[util.FullPath]struct{}
}

func (c *commandVolumeFsck) Name() string {
	return "volume.fsck"
}

func (c *commandVolumeFsck) Help() string {
	return `check all volumes to find entries not used by the filer. It is optional and resource intensive.

	Important assumption!!!
		the system is all used by one filer.

	This command works this way:
	1. collect all file ids from all volumes, as set A
	2. collect all file ids from the filer, as set B
	3. find out the set A subtract B

	If -findMissingChunksInFiler is enabled, this works
	in a reverse way:
	1. collect all file ids from all volumes, as set A
	2. collect all file ids from the filer, as set B
	3. find out the set B subtract A

	-cutoffTimeAgo is used to only check chunks older than the cutoff time.
	This is important because:
		Chunks are uploaded to volume servers before metadata is committed to filer.
		A newly uploaded chunk may appear as orphan if metadata commit is still pending.
		The default 5h cutoff provides sufficient buffer for metadata commits.

	Step 2 reads the filer from one consistent snapshot when its store supports
	that (the SQL stores do), so a file renamed while the filer is being read is
	still counted and its chunks are not mistaken for orphans. Other stores are
	walked directory by directory; there, avoid running with -reallyDeleteFromVolume
	while files are being renamed or moved.

`
}

func (c *commandVolumeFsck) HasTag(tag CommandTag) bool {
	return tag == ResourceHeavy
}

func (c *commandVolumeFsck) Do(args []string, commandEnv *CommandEnv, writer io.Writer) (err error) {

	fsckCommand := flag.NewFlagSet(c.Name(), flag.ContinueOnError)
	c.verbose = fsckCommand.Bool("v", false, "verbose mode")
	c.skipEcVolumes = fsckCommand.Bool("skipEcVolumes", false, "skip erasure coded volumes")
	c.findMissingChunksInFiler = fsckCommand.Bool("findMissingChunksInFiler", false, "see \"help volume.fsck\"")
	c.collection = fsckCommand.String("collection", "", "comma-separated collection names, wildcards, or regex patterns; empty or \"*\" matches all, \""+CollectionDefault+"\" the collection with no name")
	volumeIds := fsckCommand.String("volumeId", "", "comma separated the volume id")
	applyPurging := fsckCommand.Bool("reallyDeleteFromVolume", false, "<expert only!> after detection, delete missing data from volumes / delete missing file entries from filer. Currently this only works with default filerGroup.")
	c.forcePurging = fsckCommand.Bool("forcePurging", false, "delete missing data from volumes in one replica used together with applyPurging")
	purgeAbsent := fsckCommand.Bool("reallyDeleteFilerEntries", false, "<expert only!> delete missing file entries from filer if the corresponding volume is missing for any reason, please ensure all still existing/expected volumes are connected! used together with findMissingChunksInFiler")
	c.purgeConcurrency = fsckCommand.Int("purgeConcurrency", 16, "parallel filer deletes for -reallyDeleteFilerEntries")
	tempPath := fsckCommand.String("tempPath", path.Join(os.TempDir()), "path for temporary idx files")
	cutoffTimeAgo := fsckCommand.Duration("cutoffTimeAgo", 5*time.Hour, "only include entries on volume servers before this cutoff time to check orphan chunks")
	modifyTimeAgo := fsckCommand.Duration("modifyTimeAgo", 0, "only include entries after this modify time to check orphan chunks")
	c.verifyNeedle = fsckCommand.Bool("verifyNeedles", false, "check needles status from volume server")

	if err = fsckCommand.Parse(args); err != nil {
		return nil
	}

	// The command struct is a singleton registered in init(), so any state
	// not bound to a flag persists across shell invocations. Reset the
	// unresolved-manifest counter so a previous failed run can't permanently
	// suppress -reallyDeleteFromVolume in this session.
	c.unresolvedManifestEntries.Store(0)
	c.purgedDirs = make(map[util.FullPath]struct{})

	if err = commandEnv.confirmIsLocked(args); err != nil {
		return
	}
	c.volumeIds = make(map[uint32]bool)
	if *volumeIds != "" {
		for _, volumeIdStr := range strings.Split(*volumeIds, ",") {
			volumeIdStr = strings.TrimSpace(volumeIdStr)
			if volumeIdInt, err := strconv.ParseUint(volumeIdStr, 10, 32); err == nil {
				c.volumeIds[uint32(volumeIdInt)] = true
			} else {
				return fmt.Errorf("parse volumeId string %s to int: %v", volumeIdStr, err)
			}
		}
	}
	c.env = commandEnv
	c.writer = &lockedWriter{w: writer}

	if c.collectionMatcher, err = wildcard.CompileCollectionMatcher(*c.collection); err != nil {
		return fmt.Errorf("invalid collection pattern %q: %w", *c.collection, err)
	}

	c.bucketsPath, err = readFilerBucketsPath(commandEnv)
	if err != nil {
		return fmt.Errorf("read filer buckets path: %w", err)
	}

	// create a temp folder
	c.tempFolder, err = os.MkdirTemp(util.ResolvePath(*tempPath), "sw_fsck")
	if err != nil {
		return fmt.Errorf("failed to create temp folder: %w", err)
	}
	if *c.verbose {
		fmt.Fprintf(c.writer, "working directory: %s\n", c.tempFolder)
	}
	defer os.RemoveAll(c.tempFolder)

	c.filerSigningKey = util.GetViper().GetString("jwt.filer_signing.key")

	// collect all volume id locations
	dataNodeVolumeIdToVInfo, err := c.collectVolumeIds()
	if err != nil {
		return fmt.Errorf("failed to collect all volume locations: %w", err)
	}

	// Every volume the master knows about, before any filtering: a filer
	// entry pointing at a volume that was filtered out of this run is
	// excluded, not missing — it must never be reported or purged.
	c.knownVolumeIds = make(map[uint32]bool)
	for _, vidMap := range dataNodeVolumeIdToVInfo {
		for vid := range vidMap {
			c.knownVolumeIds[vid] = true
		}
	}

	// If the operator specified -volumeId, reject unknown ids up front instead
	// of silently filtering them out and reporting "no orphan data". A silent
	// success on a missing volume looks identical to a clean volume and hides
	// typos, already-deleted volumes, and stale scripts.
	if len(c.volumeIds) > 0 {
		known := make(map[uint32]bool)
		for _, vidMap := range dataNodeVolumeIdToVInfo {
			for vid := range vidMap {
				known[vid] = true
			}
		}
		var missing []uint32
		for vid := range c.volumeIds {
			if !known[vid] {
				missing = append(missing, vid)
			}
		}
		if len(missing) > 0 {
			sort.Slice(missing, func(i, j int) bool { return missing[i] < missing[j] })
			return fmt.Errorf("volume(s) not found on master: %v", missing)
		}
	}

	c.scopedFilerPath = c.resolveScopedFilerPath(dataNodeVolumeIdToVInfo)
	if *c.verbose && c.scopedFilerPath != "/" {
		fmt.Fprintf(c.writer, "scoping filer walk to %s\n", c.scopedFilerPath)
	}

	var collectCutoffFromAtNs int64 = 0
	if cutoffTimeAgo.Seconds() != 0 {
		collectCutoffFromAtNs = time.Now().Add(-*cutoffTimeAgo).UnixNano()
	}
	var collectModifyFromAtNs int64 = 0
	if modifyTimeAgo.Seconds() != 0 {
		collectModifyFromAtNs = time.Now().Add(-*modifyTimeAgo).UnixNano()
	}
	// collect each volume file ids
	eg, _ := errgroup.WithContext(context.Background())
	for _dataNodeId, _volumeIdToVInfo := range dataNodeVolumeIdToVInfo {
		dataNodeId, volumeIdToVInfo := _dataNodeId, _volumeIdToVInfo
		eg.Go(func() error {
			for volumeId, vinfo := range volumeIdToVInfo {
				if *c.skipEcVolumes && vinfo.isEcVolume {
					delete(volumeIdToVInfo, volumeId)
					continue
				}
				if len(c.volumeIds) > 0 {
					if _, ok := c.volumeIds[volumeId]; !ok {
						delete(volumeIdToVInfo, volumeId)
						continue
					}
				}
				if !c.collectionMatcher.Matches(vinfo.collection) {
					delete(volumeIdToVInfo, volumeId)
					continue
				}
				err = c.collectOneVolumeFileIds(dataNodeId, volumeId, vinfo)
				if err != nil {
					return fmt.Errorf("failed to collect file ids from volume %d on %s: %v", volumeId, vinfo.server, err)
				}
			}
			if *c.verbose {
				fmt.Fprintf(c.writer, "dn %+v filtered %d volumes and locations.\n", dataNodeId, len(dataNodeVolumeIdToVInfo[dataNodeId]))
			}
			return nil
		})
	}
	err = eg.Wait()
	if err != nil {
		fmt.Fprintf(c.writer, "got error: %v", err)
		return err
	}

	if *c.findMissingChunksInFiler {
		// collect all filer file ids and paths

		var queuePurge func(purgeItem)
		var finishPurge func()
		if *applyPurging || *purgeAbsent {
			queuePurge, finishPurge = startPurgeWorkers(*c.purgeConcurrency, c.httpDelete)
		}
		err = func() error {
			if err := c.collectFilerFileIdAndPaths(dataNodeVolumeIdToVInfo, *purgeAbsent, queuePurge, collectModifyFromAtNs, collectCutoffFromAtNs); err != nil {
				return fmt.Errorf("collectFilerFileIdAndPaths: %w", err)
			}
			for dataNodeId, volumeIdToVInfo := range dataNodeVolumeIdToVInfo {
				// for each volume, check filer file ids
				if err := c.findFilerChunksMissingInVolumeServers(volumeIdToVInfo, dataNodeId, queuePurge); err != nil {
					return fmt.Errorf("findFilerChunksMissingInVolumeServers: %w", err)
				}
			}
			return nil
		}()
		if finishPurge != nil {
			finishPurge()
		}
		if err != nil {
			return err
		}
		c.purgeEmptyDirectories()
	} else {
		// collect all filer file ids
		if err = c.collectFilerFileIdAndPaths(dataNodeVolumeIdToVInfo, false, nil, 0, 0); err != nil {
			return fmt.Errorf("failed to collect file ids from filer: %w", err)
		}
		// If any entry's manifest could not be resolved, our in-use fid set
		// is missing the sub-chunks behind it. Purging orphans now would
		// delete live data referenced only via the unresolved manifest, so
		// disable -reallyDeleteFromVolume for this run and tell the operator
		// to fix the broken entries first.
		applyPurgingEffective := *applyPurging
		if unresolved := c.unresolvedManifestEntries.Load(); unresolved > 0 && applyPurgingEffective {
			fmt.Fprintf(c.writer, "WARNING: %d entry(ies) had unresolvable chunk manifests; refusing to apply -reallyDeleteFromVolume to avoid deleting live sub-chunks. Fix the entries listed above (e.g. delete or repair them) and re-run.\n",
				unresolved)
			applyPurgingEffective = false
		}
		// volume file ids subtract filer file ids
		if err = c.findExtraChunksInVolumeServers(dataNodeVolumeIdToVInfo, applyPurgingEffective, uint64(collectModifyFromAtNs), uint64(collectCutoffFromAtNs)); err != nil {
			return fmt.Errorf("findExtraChunksInVolumeServers: %w", err)
		}
	}

	return nil
}

func (c *commandVolumeFsck) collectFilerFileIdAndPaths(dataNodeVolumeIdToVInfo map[string]map[uint32]VInfo, purgeAbsent bool, queuePurge func(purgeItem), collectModifyFromAtNs int64, cutoffFromAtNs int64) error {
	if *c.verbose {
		fmt.Fprintf(c.writer, "checking each file from filer path %s...\n", c.getCollectFilerFilePath())
	}

	files := make(map[uint32]*os.File)
	for _, volumeIdToServer := range dataNodeVolumeIdToVInfo {
		for vid := range volumeIdToServer {
			if _, ok := files[vid]; ok {
				continue
			}
			dst, openErr := os.OpenFile(getFilerFileIdFile(c.tempFolder, vid), os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0644)
			if openErr != nil {
				return fmt.Errorf("failed to create file %s: %v", getFilerFileIdFile(c.tempFolder, vid), openErr)
			}
			files[vid] = dst
		}
	}
	defer func() {
		for _, f := range files {
			f.Close()
		}
	}()

	return doTraverseAndSaving(c.env, c.writer, c.getCollectFilerFilePath(), false, false,
		func(ctx context.Context, entry *filer_pb.FullEntry, outputChan chan interface{}) (err error) {
			if *c.verbose && entry.Entry.IsDirectory {
				fmt.Fprintf(c.writer, "checking directory %s\n", util.NewFullPath(entry.Dir, entry.Entry.Name))
			}
			dataChunks, manifestChunks, resolveErr := filer.ResolveChunkManifest(ctx, filer.LookupFn(c.env), entry.Entry.GetChunks(), 0, math.MaxInt64, nil)
			if resolveErr != nil {
				// Cancellation/deadline isn't manifest corruption; surface it
				// so the BFS bails out cleanly without polluting the
				// unresolved-manifest counter (which would otherwise block
				// purges and mislead the operator about the failure cause).
				if errors.Is(resolveErr, context.Canceled) || errors.Is(resolveErr, context.DeadlineExceeded) {
					return resolveErr
				}
				// A single broken manifest used to abort the whole traversal,
				// leaving the operator with no way to identify orphans without
				// first fixing the broken file. Instead, record only the
				// top-level chunk fids (data chunks plus the manifest needles
				// themselves — sub-chunks behind the unreadable manifest are
				// unknown), warn, and keep going. The unresolved counter blocks
				// any purge step downstream so we never delete a sub-chunk we
				// couldn't account for.
				fmt.Fprintf(c.writer, "WARNING: ResolveChunkManifest failed for %s: %v — recording top-level chunk fids only; purging will be disabled\n",
					util.NewFullPath(entry.Dir, entry.Entry.Name), resolveErr)
				c.unresolvedManifestEntries.Add(1)
				dataChunks = entry.Entry.GetChunks()
				manifestChunks = nil
			}
			dataChunks = append(dataChunks, manifestChunks...)
			for _, chunk := range dataChunks {
				if cutoffFromAtNs != 0 && chunk.ModifiedTsNs > cutoffFromAtNs {
					continue
				}
				if collectModifyFromAtNs != 0 && chunk.ModifiedTsNs < collectModifyFromAtNs {
					continue
				}
				select {
				case outputChan <- &Item{
					vid:        chunk.Fid.VolumeId,
					fileKey:    chunk.Fid.FileKey,
					cookie:     chunk.Fid.Cookie,
					path:       util.NewFullPath(entry.Dir, entry.Entry.Name),
					hardLinkId: entry.Entry.GetHardLinkId(),
				}:
				case <-ctx.Done():
					return ctx.Err()
				}
			}
			return nil
		},
		func(outputChan chan interface{}) error {
			buffer := make([]byte, readbufferSize)
			for item := range outputChan {
				i := item.(*Item)
				if f, ok := files[i.vid]; ok {
					util.Uint64toBytes(buffer, i.fileKey)
					util.Uint32toBytes(buffer[8:], i.cookie)
					util.Uint32toBytes(buffer[12:], uint32(len(i.path)))
					if _, err := f.Write(buffer); err != nil {
						return err
					}
					if _, err := f.Write([]byte(i.path)); err != nil {
						return err
					}
					util.Uint32toBytes(buffer[:4], uint32(len(i.hardLinkId)))
					if _, err := f.Write(buffer[:4]); err != nil {
						return err
					}
					if _, err := f.Write(i.hardLinkId); err != nil {
						return err
					}
				} else if *c.findMissingChunksInFiler {
					// check if the volume matches the filter
					if len(c.volumeIds) > 0 {
						if _, ok := c.volumeIds[i.vid]; !ok {
							continue
						}
					}
					// A volume the master knows but this run filtered out is not a
					// missing volume: its files are healthy and stay untouched.
					if c.knownVolumeIds[i.vid] {
						continue
					}
					fmt.Fprintf(c.writer, "%d,%x%08x %s volume not found\n", i.vid, i.fileKey, i.cookie, i.path)
					if purgeAbsent {
						fmt.Fprintf(c.writer, "deleting path %s after volume not found\n", i.path)
						queuePurge(purgeItem{path: i.path, hardLinkId: i.hardLinkId})
					}
				}
			}
			return nil
		})
}

func (c *commandVolumeFsck) findFilerChunksMissingInVolumeServers(volumeIdToVInfo map[uint32]VInfo, dataNodeId string, queuePurge func(purgeItem)) error {

	for volumeId, vinfo := range volumeIdToVInfo {
		checkErr := c.oneVolumeFileIdsCheckOneVolume(dataNodeId, volumeId, queuePurge)
		if checkErr != nil {
			return fmt.Errorf("failed to collect file ids from volume %d on %s: %v", volumeId, vinfo.server, checkErr)
		}
	}
	return nil
}

func (c *commandVolumeFsck) findExtraChunksInVolumeServers(dataNodeVolumeIdToVInfo map[string]map[uint32]VInfo, applyPurging bool, modifyFrom, cutoffFrom uint64) error {

	var totalInUseCount, totalOrphanChunkCount, totalOrphanDataSize uint64
	// map[volumeId]map[fid]replicaCount — counts how many replicas reported
	// this fid as orphan. A fid is safe to purge without -forcePurging only
	// when replicaCount == volumeReplicaCounts[volumeId] (i.e. every replica
	// agrees it's orphan). The previous bool-based tracking treated "seen on
	// any 2 replicas" as "seen on all replicas", which was wrong for
	// 3+-replica volumes.
	volumeIdOrphanFileIds := make(map[uint32]map[string]int)
	volumeReplicaCounts := make(map[uint32]int)
	isEcVolumeReplicas := make(map[uint32]bool)
	// Track which specific replicas were read-only so we only flip those
	// back on exit. The old `isReadOnlyReplicas[volumeId] = bool` leaked
	// read-only state across replicas: if one replica was RO and another RW,
	// the deferred cleanup would mark the originally-RW replica RO too.
	readOnlyServerReplicas := make(map[uint32][]pb.ServerAddress)
	// Phase 1: collect orphan fids from every replica of every volume.
	// The purge step runs in Phase 2, AFTER every replica has contributed.
	// Running purge inside this loop (as the original code did) meant the
	// first replica's orphans were deleted before later replicas could
	// participate in the intersection — so the "only purge fids seen on
	// all replicas" safety net only worked by accident, and purge also
	// fired multiple times per volume (once per data-node iteration).
	for dataNodeId, volumeIdToVInfo := range dataNodeVolumeIdToVInfo {
		for volumeId, vinfo := range volumeIdToVInfo {
			inUseCount, orphanFileIds, orphanDataSize, checkErr := c.oneVolumeFileIdsSubtractFilerFileIds(dataNodeId, volumeId, &vinfo, modifyFrom, cutoffFrom)
			if checkErr != nil {
				return fmt.Errorf("failed to collect file ids from volume %d on %s: %v", volumeId, vinfo.server, checkErr)
			}
			if _, found := volumeIdOrphanFileIds[volumeId]; !found {
				volumeIdOrphanFileIds[volumeId] = make(map[string]int)
			}
			volumeReplicaCounts[volumeId]++
			for _, fid := range orphanFileIds {
				volumeIdOrphanFileIds[volumeId][fid]++
			}

			totalInUseCount += inUseCount
			totalOrphanChunkCount += uint64(len(orphanFileIds))
			totalOrphanDataSize += orphanDataSize

			if *c.verbose {
				for _, fid := range orphanFileIds {
					fmt.Fprintf(c.writer, "%s:%s\n", vinfo.collection, fid)
				}
			}
			isEcVolumeReplicas[volumeId] = vinfo.isEcVolume
			if vinfo.isReadOnly {
				readOnlyServerReplicas[volumeId] = append(readOnlyServerReplicas[volumeId], vinfo.server)
			}
		}
	}

	// Phase 2: purge. At most one call to purgeFileIdsForOneVolume per
	// volume — that helper already fans out to all replica locations via
	// MasterClient.GetLocations, so iterating per replica here (as the old
	// code did) would issue N*N delete RPCs for N replicas.
	if applyPurging {
		var skippedVolumeIds []uint32
		for volumeId, orphanReplicaFileIds := range volumeIdOrphanFileIds {
			if len(orphanReplicaFileIds) == 0 {
				continue
			}
			if isEcVolumeReplicas[volumeId] {
				fmt.Fprintf(c.writer, "skip purging for Erasure Coded volume %d.\n", volumeId)
				continue
			}
			// Call out to a closure per volume so the deferred "mark
			// readonly again" fires between volumes instead of piling up
			// until findExtraChunksInVolumeServers returns. Per-volume
			// failures (e.g. a replica stuck read-only) don't halt the
			// rest of the run; the volume is remembered and its deletes
			// are skipped.
			if err := c.purgeOneVolume(volumeId, orphanReplicaFileIds, volumeReplicaCounts[volumeId], readOnlyServerReplicas[volumeId]); err != nil {
				fmt.Fprintf(c.writer, "skip purging volume %d: %v\n", volumeId, err)
				skippedVolumeIds = append(skippedVolumeIds, volumeId)
			}
		}
		if len(skippedVolumeIds) > 0 {
			sort.Slice(skippedVolumeIds, func(i, j int) bool { return skippedVolumeIds[i] < skippedVolumeIds[j] })
			fmt.Fprintf(c.writer, "skipped purge on %d volume(s): %v\n", len(skippedVolumeIds), skippedVolumeIds)
		}
	}

	if !applyPurging {
		var pct float64

		if totalCount := totalOrphanChunkCount + totalInUseCount; totalCount > 0 {
			pct = float64(totalOrphanChunkCount) * 100 / (float64(totalCount))
		}

		fmt.Fprintf(c.writer, "\nTotal\t\tentries:%d\torphan:%d\t%.2f%%\t%dB\n",
			totalOrphanChunkCount+totalInUseCount, totalOrphanChunkCount, pct, totalOrphanDataSize)

		fmt.Fprintf(c.writer, "This could be normal if multiple filers or no filers are used.\n")
	}

	if totalOrphanChunkCount == 0 {
		fmt.Fprintf(c.writer, "no orphan data\n")
	}

	return nil
}

// purgeOneVolume picks the orphan fids to delete for a single volume and
// fires the delete RPC. It's split out of findExtraChunksInVolumeServers so
// the `defer markVolumeWritable(context.Background(), ..., false, false)` at the bottom fires
// between volumes — putting that defer inside the caller's for-loop would
// leave every processed volume writable until the whole fsck run finished.
func (c *commandVolumeFsck) purgeOneVolume(volumeId uint32, orphanReplicaFileIds map[string]int, replicaCount int, readOnlyReplicas []pb.ServerAddress) error {
	orphanFileIds := make([]string, 0, len(orphanReplicaFileIds))
	for fid, foundInReplicaCount := range orphanReplicaFileIds {
		// Default safety net: only purge fids every replica reported as
		// orphan. -forcePurging bypasses this for operators who've already
		// decided they're OK with the single-replica evidence.
		if foundInReplicaCount == replicaCount || *c.forcePurging {
			orphanFileIds = append(orphanFileIds, fid)
		}
	}
	if len(orphanFileIds) == 0 {
		return nil
	}
	if *c.verbose {
		fmt.Fprintf(c.writer, "purging process for volume %d.\n", volumeId)
	}

	needleVID := needle.VolumeId(volumeId)
	for _, server := range readOnlyReplicas {
		if err := markVolumeWritable(context.Background(), c.env.option.GrpcDialOption, needleVID, server, true, false); err != nil {
			// Replicas flipped writable earlier roll back via the defer.
			return fmt.Errorf("mark %v writable: %v", server, err)
		}
		fmt.Fprintf(c.writer, "temporarily marked %d on server %v writable for forced purge\n", volumeId, server)
		defer markVolumeWritable(context.Background(), c.env.option.GrpcDialOption, needleVID, server, false, false)
	}

	if *c.verbose {
		fmt.Fprintf(c.writer, "purging files from volume %d\n", volumeId)
	}

	if err := c.purgeFileIdsForOneVolume(volumeId, orphanFileIds); err != nil {
		return fmt.Errorf("purging volume %d: %v", volumeId, err)
	}
	return nil
}

func (c *commandVolumeFsck) collectOneVolumeFileIds(dataNodeId string, volumeId uint32, vinfo VInfo) error {

	if *c.verbose {
		fmt.Fprintf(c.writer, "collecting volume %d file ids from %s ...\n", volumeId, vinfo.server)
	}

	return operation.WithVolumeServerClient(false, vinfo.server, c.env.option.GrpcDialOption,
		func(volumeServerClient volume_server_pb.VolumeServerClient) error {
			ext := ".idx"
			if vinfo.isEcVolume {
				ext = ".ecx"
			}

			copyFileClient, err := volumeServerClient.CopyFile(context.Background(), &volume_server_pb.CopyFileRequest{
				VolumeId:                 volumeId,
				Ext:                      ext,
				CompactionRevision:       math.MaxUint32,
				StopOffset:               math.MaxInt64,
				Collection:               vinfo.collection,
				IsEcVolume:               vinfo.isEcVolume,
				IgnoreSourceFileNotFound: false,
			})
			if err != nil {
				return fmt.Errorf("failed to start copying volume %d%s: %v", volumeId, ext, err)
			}

			var buf bytes.Buffer
			for {
				resp, err := copyFileClient.Recv()
				if errors.Is(err, io.EOF) {
					break
				}
				if err != nil {
					return err
				}
				buf.Write(resp.FileContent)
			}
			idxFilename := getVolumeFileIdFile(c.tempFolder, dataNodeId, volumeId)
			err = writeToFile(buf.Bytes(), idxFilename)
			if err != nil {
				return fmt.Errorf("failed to copy %d%s from %s: %v", volumeId, ext, vinfo.server, err)
			}

			return nil
		})

}

type Item struct {
	vid        uint32
	fileKey    uint64
	cookie     uint32
	path       util.FullPath
	hardLinkId []byte
}

// lockedWriter serializes concurrent writers — the traversal, the consumer,
// and the purge workers all log through c.writer at the same time.
type lockedWriter struct {
	mu sync.Mutex
	w  io.Writer
}

func (lw *lockedWriter) Write(p []byte) (int, error) {
	lw.mu.Lock()
	defer lw.mu.Unlock()
	return lw.w.Write(p)
}

func (c *commandVolumeFsck) readFilerFileIdFile(volumeId uint32, fn func(needleId types.NeedleId, itemPath util.FullPath, hardLinkId []byte)) error {
	fp, err := os.Open(getFilerFileIdFile(c.tempFolder, volumeId))
	if err != nil {
		return err
	}
	defer fp.Close()

	br := bufio.NewReader(fp)
	buffer := make([]byte, readbufferSize)
	var readSize int
	var readErr error
	item := &Item{vid: volumeId}
	for {
		readSize, readErr = io.ReadFull(br, buffer)
		if errors.Is(readErr, io.EOF) {
			break
		}
		if readErr != nil {
			return fmt.Errorf("read fid header for volume %d: %w", volumeId, readErr)
		}
		if readSize != readbufferSize {
			return fmt.Errorf("read fid header size mismatch for volume %d: got %d want %d", volumeId, readSize, readbufferSize)
		}
		item.fileKey = util.BytesToUint64(buffer[:8])
		item.cookie = util.BytesToUint32(buffer[8:12])
		pathSize := util.BytesToUint32(buffer[12:16])
		pathBytes := make([]byte, int(pathSize))
		n, err := io.ReadFull(br, pathBytes)
		if err != nil {
			return fmt.Errorf("read fid path for volume %d,%x%08x: %w", volumeId, item.fileKey, item.cookie, err)
		}
		if n != int(pathSize) {
			return fmt.Errorf("read fid path size mismatch for volume %d,%x%08x: got %d want %d", volumeId, item.fileKey, item.cookie, n, pathSize)
		}
		item.path = util.FullPath(pathBytes)
		item.hardLinkId = nil
		if _, err := io.ReadFull(br, buffer[:4]); err != nil {
			return fmt.Errorf("read fid hard link id size for volume %d,%x%08x: %w", volumeId, item.fileKey, item.cookie, err)
		}
		if hardLinkIdSize := util.BytesToUint32(buffer[:4]); hardLinkIdSize > 0 {
			item.hardLinkId = make([]byte, int(hardLinkIdSize))
			if _, err := io.ReadFull(br, item.hardLinkId); err != nil {
				return fmt.Errorf("read fid hard link id for volume %d,%x%08x: %w", volumeId, item.fileKey, item.cookie, err)
			}
		}
		needleId := types.NeedleId(item.fileKey)
		fn(needleId, item.path, item.hardLinkId)
	}
	return nil
}

func (c *commandVolumeFsck) oneVolumeFileIdsCheckOneVolume(dataNodeId string, volumeId uint32, queuePurge func(purgeItem)) (err error) {
	if *c.verbose {
		fmt.Fprintf(c.writer, "find missing file chunks in dataNodeId %s volume %d ...\n", dataNodeId, volumeId)
	}

	db := needle_map.NewMemDb()
	defer db.Close()

	if err = db.LoadFromIdx(getVolumeFileIdFile(c.tempFolder, dataNodeId, volumeId)); err != nil {
		return
	}
	if err = c.readFilerFileIdFile(volumeId, func(needleId types.NeedleId, itemPath util.FullPath, hardLinkId []byte) {
		if _, found := db.Get(needleId); !found {
			fmt.Fprintf(c.writer, "%s\n", itemPath)
			if queuePurge != nil {
				queuePurge(purgeItem{path: itemPath, hardLinkId: hardLinkId})
			}
		}
	}); err != nil {
		return
	}
	return nil
}

type purgeItem struct {
	path       util.FullPath
	hardLinkId []byte
}

// startPurgeWorkers fans deletes across at least one worker and returns a
// queue function plus a finish function that closes the channels and waits
// for in-flight deletes. Paths that share a hard link id always land on the
// same worker, so their shared hard-link counter updates stay serialized —
// concurrent deletes of the same id could both read count N and write N-1,
// leaving the shared record behind after every path is gone.
func startPurgeWorkers(workers int, deleteFn func(util.FullPath)) (queue func(purgeItem), finish func()) {
	if workers < 1 {
		workers = 1
	}
	chans := make([]chan purgeItem, workers)
	var wg sync.WaitGroup
	for w := range chans {
		chans[w] = make(chan purgeItem, 128)
		wg.Add(1)
		go func(ch <-chan purgeItem) {
			defer wg.Done()
			for it := range ch {
				deleteFn(it.path)
			}
		}(chans[w])
	}
	return func(it purgeItem) {
			key := []byte(it.path)
			if len(it.hardLinkId) > 0 {
				key = it.hardLinkId
			}
			chans[int(uint32(util.HashToInt32(key))%uint32(workers))] <- it
		}, func() {
			for _, ch := range chans {
				close(ch)
			}
			wg.Wait()
		}
}

func (c *commandVolumeFsck) httpDelete(path util.FullPath) {
	req, err := http.NewRequest(http.MethodDelete, "", nil)
	if err != nil {
		fmt.Fprintf(c.writer, "HTTP delete request error: %v\n", err)
		return
	}

	req.URL = &url.URL{
		Scheme: "http",
		Host:   c.env.option.FilerAddress.ToHttpAddress(),
		Path:   string(path),
	}

	if c.filerSigningKey != "" {
		encodedJwt := security.GenJwtForFilerServer(security.SigningKey(c.filerSigningKey), jwtFilerTokenExpirationSeconds)
		req.Header.Set("Authorization", security.BearerPrefix+string(encodedJwt))
	}

	if *c.verbose {
		fmt.Fprintf(c.writer, "full HTTP delete request to be sent: %v\n", req)
	}

	resp, err := util_http.GetGlobalHttpClient().Do(req)
	if err != nil {
		fmt.Fprintf(c.writer, "DELETE fetch error: %v\n", err)
		return
	}
	defer resp.Body.Close()

	_, err = io.ReadAll(resp.Body)
	if err != nil {
		fmt.Fprintf(c.writer, "DELETE response error: %v\n", err)
	}

	if *c.verbose {
		fmt.Fprintln(c.writer, "delete response Status : ", resp.Status)
		fmt.Fprintln(c.writer, "delete response Headers : ", resp.Header)
	}

	if resp.StatusCode < http.StatusBadRequest {
		dir, _ := path.DirAndName()
		c.purgedDirsLock.Lock()
		c.purgedDirs[util.FullPath(dir)] = struct{}{}
		c.purgedDirsLock.Unlock()
	}
}

// purgeEmptyDirectories removes the directories emptied by the purged entries, walking up while each parent is empty too.
func (c *commandVolumeFsck) purgeEmptyDirectories() {
	candidates := make(map[util.FullPath]struct{})
	c.purgedDirsLock.Lock()
	for dir := range c.purgedDirs {
		for d := dir; c.canPurgeDirectory(d); {
			candidates[d] = struct{}{}
			parent, _ := d.DirAndName()
			d = util.FullPath(parent)
		}
	}
	c.purgedDirsLock.Unlock()

	dirs := make([]util.FullPath, 0, len(candidates))
	for dir := range candidates {
		dirs = append(dirs, dir)
	}
	// deepest first, so a directory is only tried once its children are gone
	sort.Slice(dirs, func(i, j int) bool { return len(dirs[i]) > len(dirs[j]) })

	for _, dir := range dirs {
		entry, _, _, lookupErr := filer_pb.GetEntry(context.Background(), c.env, dir)
		if lookupErr != nil && !errors.Is(lookupErr, filer_pb.ErrNotFound) {
			fmt.Fprintf(c.writer, "lookup directory %s: %v\n", dir, lookupErr)
			continue
		}
		// a directory key object is an S3 object of its own
		if entry == nil || entry.IsDirectoryKeyObject() {
			continue
		}
		// a zero mtime turns the delete's condition off, leaving nothing to hold it to
		mtime := entry.Attributes.GetMtime()
		if mtime <= 0 || mtime >= time.Now().Add(-directoryQuietPeriod).Unix() {
			continue
		}
		if err := c.deleteEmptyDirectory(dir, mtime); err != nil {
			if !errors.Is(err, filer.ErrNonEmptyFolder) {
				fmt.Fprintf(c.writer, "delete empty directory %s: %v\n", dir, err)
			}
			continue
		}
		fmt.Fprintf(c.writer, "deleted empty directory %s\n", dir)
	}
}

// deleteEmptyDirectory deletes dir unless it changed since it was looked up at mtime,
// so a directory promoted to an S3 object meanwhile survives. The delete is not
// recursive, leaving the filer to reject a directory that is not empty.
func (c *commandVolumeFsck) deleteEmptyDirectory(dir util.FullPath, mtime int64) error {
	parent, name := dir.DirAndName()
	return c.env.WithFilerClient(false, func(client filer_pb.SeaweedFilerClient) error {
		resp, err := client.DeleteEntry(context.Background(), &filer_pb.DeleteEntryRequest{
			Directory:          parent,
			Name:               name,
			IfNotModifiedAfter: mtime,
		})
		if err != nil {
			return err
		}
		if resp.Error != "" {
			return filer.DeleteEntryError(resp.Error)
		}
		return nil
	})
}

func (c *commandVolumeFsck) canPurgeDirectory(dir util.FullPath) bool {
	root := c.getCollectFilerFilePath()
	if string(dir) == root || !strings.HasPrefix(string(dir), strings.TrimSuffix(root, "/")+"/") {
		return false
	}
	// deleting a bucket drops its whole collection
	parent, _ := dir.DirAndName()
	return string(dir) != c.bucketsPath && parent != c.bucketsPath
}

func (c *commandVolumeFsck) oneVolumeFileIdsSubtractFilerFileIds(dataNodeId string, volumeId uint32, vinfo *VInfo, modifyFrom, cutoffFrom uint64) (inUseCount uint64, orphanFileIds []string, orphanDataSize uint64, err error) {

	volumeFileIdDb := needle_map.NewMemDb()
	defer volumeFileIdDb.Close()

	if err = volumeFileIdDb.LoadFromIdx(getVolumeFileIdFile(c.tempFolder, dataNodeId, volumeId)); err != nil {
		err = fmt.Errorf("failed to LoadFromIdx %+v", err)
		return
	}

	if err = c.readFilerFileIdFile(volumeId, func(filerNeedleId types.NeedleId, itemPath util.FullPath, _ []byte) {
		inUseCount++
		if *c.verifyNeedle && !vinfo.isEcVolume {
			if needleValue, ok := volumeFileIdDb.Get(filerNeedleId); ok && !needleValue.Size.IsDeleted() {
				if _, err := readNeedleStatus(c.env.option.GrpcDialOption, vinfo.server, volumeId, *needleValue); err != nil {
					// files may be deleted during copying filesIds
					if !strings.Contains(err.Error(), storage.ErrorDeleted.Error()) {
						fmt.Fprintf(c.writer, "failed to read %d:%s needle status of file %s: %+v\n",
							volumeId, filerNeedleId.String(), itemPath, err)
						if *c.forcePurging {
							return
						}
					}
				}
			}
		}

		if err = volumeFileIdDb.Delete(filerNeedleId); err != nil && *c.verbose {
			fmt.Fprintf(c.writer, "failed to nm.delete %s(%+v): %+v", itemPath, filerNeedleId, err)
		}
	}); err != nil {
		err = fmt.Errorf("failed to readFilerFileIdFile %+v", err)
		return
	}

	var orphanFileCount uint64
	if err = volumeFileIdDb.AscendingVisit(func(n needle_map.NeedleValue) error {
		if n.Size.IsDeleted() {
			return nil
		}
		if !vinfo.isEcVolume && (cutoffFrom > 0 || modifyFrom > 0) {
			return operation.WithVolumeServerClient(false, vinfo.server, c.env.option.GrpcDialOption,
				func(volumeServerClient volume_server_pb.VolumeServerClient) error {
					resp, err := volumeServerClient.ReadNeedleMeta(context.Background(), &volume_server_pb.ReadNeedleMetaRequest{
						VolumeId: volumeId,
						NeedleId: types.NeedleIdToUint64(n.Key),
						Offset:   n.Offset.ToActualOffset(),
						Size:     int32(n.Size),
					})
					if err != nil {
						return fmt.Errorf("read needle meta with id %d from volume %d: %v", n.Key, volumeId, err)
					}
					if (modifyFrom == 0 || modifyFrom <= resp.AppendAtNs) && (cutoffFrom == 0 || resp.AppendAtNs <= cutoffFrom) {
						orphanFileIds = append(orphanFileIds, n.Key.FileId(volumeId))
						orphanFileCount++
						orphanDataSize += uint64(n.Size)
					}
					return nil
				})
		} else {
			if vinfo.isEcVolume && (cutoffFrom > 0 || modifyFrom > 0) {
				if *c.verbose {
					fmt.Fprintf(c.writer, "skipping time-based filtering for EC volume %d (cutoffFrom=%d, modifyFrom=%d)\n", volumeId, cutoffFrom, modifyFrom)
				}
			}
			orphanFileIds = append(orphanFileIds, n.Key.FileId(volumeId))
			orphanFileCount++
			orphanDataSize += uint64(n.Size)
		}
		return nil
	}); err != nil {
		err = fmt.Errorf("failed to AscendingVisit %+v", err)
		return
	}

	if orphanFileCount > 0 {
		pct := float64(orphanFileCount*100) / (float64(orphanFileCount + inUseCount))
		fmt.Fprintf(c.writer, "dataNode:%s\tvolume:%d\tentries:%d\torphan:%d\t%.2f%%\t%dB\n",
			dataNodeId, volumeId, orphanFileCount+inUseCount, orphanFileCount, pct, orphanDataSize)
	}

	return

}

type VInfo struct {
	server     pb.ServerAddress
	collection string
	isEcVolume bool
	isReadOnly bool
}

func (c *commandVolumeFsck) collectVolumeIds() (volumeIdToServer map[string]map[uint32]VInfo, err error) {

	if *c.verbose {
		fmt.Fprintf(c.writer, "collecting volume id and locations from master ...\n")
	}

	volumeIdToServer = make(map[string]map[uint32]VInfo)
	// collect topology information
	topologyInfo, _, err := collectTopologyInfo(c.env, 0)
	if err != nil {
		return
	}

	eachDataNode(topologyInfo, func(dc DataCenterId, rack RackId, t *master_pb.DataNodeInfo) {
		var volumeCount, ecShardCount int
		dataNodeId := t.GetId()
		for _, diskInfo := range t.DiskInfos {
			if _, ok := volumeIdToServer[dataNodeId]; !ok {
				volumeIdToServer[dataNodeId] = make(map[uint32]VInfo)
			}
			for _, vi := range diskInfo.VolumeInfos {
				volumeIdToServer[dataNodeId][vi.Id] = VInfo{
					server:     pb.NewServerAddressFromDataNode(t),
					collection: vi.Collection,
					isEcVolume: false,
					isReadOnly: vi.ReadOnly,
				}
				volumeCount += 1
			}
			for _, ecShardInfo := range diskInfo.EcShardInfos {
				volumeIdToServer[dataNodeId][ecShardInfo.Id] = VInfo{
					server:     pb.NewServerAddressFromDataNode(t),
					collection: ecShardInfo.Collection,
					isEcVolume: true,
					isReadOnly: true,
				}
				ecShardCount += 1
			}
		}
		if *c.verbose {
			fmt.Fprintf(c.writer, "dn %+v collected %d volumes and %d ec shards.\n", dataNodeId, volumeCount, ecShardCount)
		}
	})
	return
}

func (c *commandVolumeFsck) purgeFileIdsForOneVolume(volumeId uint32, fileIds []string) (err error) {
	fmt.Fprintf(c.writer, "purging orphan data for volume %d...\n", volumeId)
	locations, found := c.env.MasterClient.GetLocations(volumeId)
	if !found {
		return fmt.Errorf("failed to find volume %d locations", volumeId)
	}

	resultChan := make(chan []*volume_server_pb.DeleteResult, len(locations))
	var wg sync.WaitGroup
	for _, location := range locations {
		wg.Add(1)
		go func(server pb.ServerAddress, fidList []string) {
			defer wg.Done()

			deleteResults := operation.DeleteFileIdsAtOneVolumeServer(context.Background(), server, c.env.option.GrpcDialOption, fidList, false)
			if deleteResults != nil {
				resultChan <- deleteResults
			}

		}(location.ServerAddress(), fileIds)
	}
	wg.Wait()
	close(resultChan)

	for results := range resultChan {
		for _, result := range results {
			if result.Error != "" {
				fmt.Fprintf(c.writer, "purge error: %s\n", result.Error)
			}
		}
	}

	return
}

func (c *commandVolumeFsck) getCollectFilerFilePath() string {
	return c.scopedFilerPath
}

func (c *commandVolumeFsck) resolveScopedFilerPath(dataNodeVolumeIdToVInfo map[string]map[uint32]VInfo) string {
	if *c.collection != "" {
		// fs.configure can bind a collection to any path prefix, so files
		// referencing the filtered volumes may live anywhere. A bucket-dir
		// walk could miss them and report their live chunks as orphans;
		// only a full walk is safe for a collection filter.
		return "/"
	}
	if len(c.volumeIds) == 0 {
		return "/"
	}
	collections := make(map[string]struct{})
	for _, vidMap := range dataNodeVolumeIdToVInfo {
		for vid, vinfo := range vidMap {
			if _, ok := c.volumeIds[vid]; !ok {
				continue
			}
			// empty collection: volume can be referenced from anywhere
			if vinfo.collection == "" {
				return "/"
			}
			collections[vinfo.collection] = struct{}{}
			if len(collections) > 1 {
				return "/"
			}
		}
	}
	if len(collections) != 1 {
		return "/"
	}
	var collection string
	for col := range collections {
		collection = col
	}
	exists, err := c.bucketDirExists(collection)
	if err != nil || !exists {
		return "/"
	}
	return fmt.Sprintf("%s/%s", c.bucketsPath, collection)
}

func (c *commandVolumeFsck) bucketDirExists(name string) (bool, error) {
	var found bool
	err := c.env.WithFilerClient(false, func(client filer_pb.SeaweedFilerClient) error {
		resp, err := client.LookupDirectoryEntry(context.Background(), &filer_pb.LookupDirectoryEntryRequest{
			Directory: c.bucketsPath,
			Name:      name,
		})
		if err != nil {
			if strings.Contains(err.Error(), filer_pb.ErrNotFound.Error()) {
				return nil
			}
			return err
		}
		if resp.Entry != nil && resp.Entry.IsDirectory {
			found = true
		}
		return nil
	})
	return found, err
}

func getVolumeFileIdFile(tempFolder string, dataNodeid string, vid uint32) string {
	return filepath.Join(tempFolder, fmt.Sprintf("%s_%d.idx", dataNodeid, vid))
}

func getFilerFileIdFile(tempFolder string, vid uint32) string {
	return filepath.Join(tempFolder, fmt.Sprintf("%d.fid", vid))
}

func writeToFile(bytes []byte, fileName string) error {
	flags := os.O_WRONLY | os.O_CREATE | os.O_TRUNC
	dst, err := os.OpenFile(fileName, flags, 0644)
	if err != nil {
		return err
	}
	defer dst.Close()

	_, err = dst.Write(bytes)
	return err
}