Files
seaweedfs/weed/mount/weedfs_file_copy_range.go
T
Chris Lu b5a952bcb1 fix(mount): don't strand a directory cached-but-empty when an eviction races a rebuild (#9791)
* fix(mount): don't strand a directory cached-but-empty when an off-loop wipe races a rebuild

Idle eviction, kernel Forget, and the copy-range fallback cleared a
directory's cached entries directly, off the metaCache apply loop, after
resetting the cached flag in inodeToPath as a separate step. A concurrent
rebuild could publish a fresh listing (markCachedFn) in between, so the late
DeleteFolderChildren left the directory flagged cached over an empty store.
lookupEntry then returns an authoritative ENOENT and ReadDir returns nothing,
so every file in the directory disappears from the mount although it is still
present on the filer.

Route those wipes through a new apply-loop step that resets the flag and wipes
the store together, serialized with a build's markCachedFn, and skips a
directory while it is building.

* fix(mount): route the meta-event retry cleanup through the apply-loop purge

The subscription-retry callback wiped the mount root's cached children
directly off the apply loop and reset the cache flags as a separate step — the
same pattern that can leave a concurrently-rebuilding root cached-but-empty.
Invalidate all flags (safe on its own, it never deletes entries) then purge the
root's children through the apply loop.
2026-06-02 14:43:46 -07:00

508 lines
18 KiB
Go

package mount
import (
"bytes"
"context"
"fmt"
"io"
"math"
"net/http"
"net/url"
"time"
"github.com/seaweedfs/go-fuse/v2/fuse"
"google.golang.org/protobuf/proto"
"github.com/seaweedfs/seaweedfs/weed/filer"
"github.com/seaweedfs/seaweedfs/weed/glog"
"github.com/seaweedfs/seaweedfs/weed/pb/filer_pb"
"github.com/seaweedfs/seaweedfs/weed/security"
"github.com/seaweedfs/seaweedfs/weed/util"
util_http "github.com/seaweedfs/seaweedfs/weed/util/http"
request_id "github.com/seaweedfs/seaweedfs/weed/util/request_id"
)
type serverSideWholeFileCopyOutcome uint8
const (
serverSideWholeFileCopyNotCommitted serverSideWholeFileCopyOutcome = iota
serverSideWholeFileCopyCommitted
serverSideWholeFileCopyAmbiguous
)
type wholeFileServerCopyRequest struct {
srcPath util.FullPath
dstPath util.FullPath
sourceSize int64
srcInode uint64
srcMtime int64
dstInode uint64
dstMtime int64
dstSize int64
sourceMime string
sourceMd5 []byte
copyRequestID string
}
// performServerSideWholeFileCopy is a package-level seam so tests can override
// the filer call without standing up an HTTP endpoint.
var performServerSideWholeFileCopy = func(cancel <-chan struct{}, wfs *WFS, copyRequest wholeFileServerCopyRequest) (*filer_pb.Entry, serverSideWholeFileCopyOutcome, error) {
return wfs.copyEntryViaFiler(cancel, copyRequest)
}
// filerCopyRequestTimeout bounds the mount->filer POST so a stalled copy does
// not block copy_file_range workers indefinitely.
const filerCopyRequestTimeout = 60 * time.Second
// filerCopyReadbackTimeout gives the follow-up metadata reload a fresh deadline
// after the filer already accepted the copy request.
const filerCopyReadbackTimeout = 15 * time.Second
// CopyFileRange copies data from one file to another from and to specified offsets.
//
// See https://man7.org/linux/man-pages/man2/copy_file_range.2.html
// See https://github.com/libfuse/libfuse/commit/fe4f9428fc403fa8b99051f52d84ea5bd13f3855
/**
* Copy a range of data from one file to another
*
* Niels de Vos: • libfuse: add copy_file_range() support
*
* Performs an optimized copy between two file descriptors without the
* additional cost of transferring data through the FUSE kernel module
* to user space (glibc) and then back into the FUSE filesystem again.
*
* In case this method is not implemented, applications are expected to
* fall back to a regular file copy. (Some glibc versions did this
* emulation automatically, but the emulation has been removed from all
* glibc release branches.)
*/
func (wfs *WFS) CopyFileRange(cancel <-chan struct{}, in *fuse.CopyFileRangeIn) (written uint32, code fuse.Status) {
// flags must equal 0 for this syscall as of now
if in.Flags != 0 {
return 0, fuse.EINVAL
}
// files must exist
fhOut := wfs.GetHandle(FileHandleId(in.FhOut))
if fhOut == nil {
return 0, fuse.EBADF
}
fhIn := wfs.GetHandle(FileHandleId(in.FhIn))
if fhIn == nil {
return 0, fuse.EBADF
}
// lock source and target file handles
fhOutActiveLock := fhOut.wfs.fhLockTable.AcquireLock("CopyFileRange", fhOut.fh, util.ExclusiveLock)
defer fhOut.wfs.fhLockTable.ReleaseLock(fhOut.fh, fhOutActiveLock)
if fhOut.entry == nil {
return 0, fuse.ENOENT
}
if fhIn.fh != fhOut.fh {
fhInActiveLock := fhIn.wfs.fhLockTable.AcquireLock("CopyFileRange", fhIn.fh, util.SharedLock)
defer fhIn.wfs.fhLockTable.ReleaseLock(fhIn.fh, fhInActiveLock)
}
// directories are not supported
if fhIn.entry.IsDirectory || fhOut.entry.IsDirectory {
return 0, fuse.EISDIR
}
glog.V(4).Infof(
"CopyFileRange %s fhIn %d -> %s fhOut %d, [%d,%d) -> [%d,%d)",
fhIn.FullPath(), fhIn.fh,
fhOut.FullPath(), fhOut.fh,
in.OffIn, in.OffIn+in.Len,
in.OffOut, in.OffOut+in.Len,
)
if written, handled, status := wfs.tryServerSideWholeFileCopy(cancel, in, fhIn, fhOut); handled {
return written, status
}
// Concurrent copy operations could allocate too much memory, so we want to
// throttle our concurrency, scaling with the number of writers the mount
// was configured with.
if wfs.concurrentCopiersSem != nil {
wfs.concurrentCopiersSem <- struct{}{}
defer func() { <-wfs.concurrentCopiersSem }()
}
// We want to stream the copy operation to avoid allocating massive buffers.
nowUnixNano := time.Now().UnixNano()
totalCopied := int64(0)
buff := wfs.copyBufferPool.Get().([]byte)
defer wfs.copyBufferPool.Put(buff)
for {
// Comply with cancellation as best as we can, given that the underlying
// IO functions aren't cancellation-aware.
select {
case <-cancel:
glog.Warningf("canceled CopyFileRange for %s (copied %d)",
fhIn.FullPath(), totalCopied)
return uint32(totalCopied), fuse.EINTR
default: // keep going
}
// We can save one IO by breaking early if we already know the next read
// will result in zero bytes.
remaining := int64(in.Len) - totalCopied
readLen := min(remaining, int64(len(buff)))
if readLen == 0 {
break
}
// Perform the read
offsetIn := totalCopied + int64(in.OffIn)
numBytesRead, err := readDataByFileHandle(
buff[:readLen], fhIn, offsetIn)
if err != nil {
glog.Warningf("file handle read %s %d (total %d): %v",
fhIn.FullPath(), numBytesRead, totalCopied, err)
return 0, fuse.EIO
}
// Break if we're done copying (no more bytes to read)
if numBytesRead == 0 {
break
}
offsetOut := int64(in.OffOut) + totalCopied
// Detect mime type only during the beginning of our stream, since
// DetectContentType is expecting some of the first 512 bytes of the
// file. See [http.DetectContentType] for details.
if offsetOut <= 512 {
fhOut.contentType = http.DetectContentType(buff[:numBytesRead])
}
// Perform the write
fhOut.dirtyPages.writerPattern.MonitorWriteAt(offsetOut, int(numBytesRead))
if err := fhOut.dirtyPages.AddPage(
offsetOut,
buff[:numBytesRead],
fhOut.dirtyPages.writerPattern.IsSequentialMode(),
nowUnixNano); err != nil {
glog.Errorf("AddPage error: %v", err)
return 0, fuse.EIO
}
// Accumulate for the next loop iteration
totalCopied += numBytesRead
}
if totalCopied == 0 {
return 0, fuse.OK
}
fhOut.entry.Attributes.FileSize = uint64(max(
totalCopied+int64(in.OffOut),
int64(fhOut.entry.Attributes.FileSize),
))
fhOut.entry.Content = nil
fhOut.dirtyMetadata = true
written = uint32(totalCopied)
return written, fuse.OK
}
func (wfs *WFS) tryServerSideWholeFileCopy(cancel <-chan struct{}, in *fuse.CopyFileRangeIn, fhIn, fhOut *FileHandle) (written uint32, handled bool, code fuse.Status) {
copyRequest, ok := wholeFileServerCopyCandidate(fhIn, fhOut, in)
if !ok {
return 0, false, fuse.OK
}
glog.V(1).Infof("CopyFileRange server-side copy %s => %s (%d bytes)", copyRequest.srcPath, copyRequest.dstPath, copyRequest.sourceSize)
entry, outcome, err := performServerSideWholeFileCopy(cancel, wfs, copyRequest)
switch outcome {
case serverSideWholeFileCopyCommitted:
if err != nil {
glog.Warningf("CopyFileRange server-side copy %s => %s committed but local refresh failed: %v", copyRequest.srcPath, copyRequest.dstPath, err)
} else {
glog.V(1).Infof("CopyFileRange server-side copy %s => %s completed (%d bytes)", copyRequest.srcPath, copyRequest.dstPath, copyRequest.sourceSize)
}
wfs.applyServerSideWholeFileCopyResult(fhIn, fhOut, copyRequest.dstPath, entry, copyRequest.sourceSize)
return uint32(copyRequest.sourceSize), true, fuse.OK
case serverSideWholeFileCopyAmbiguous:
glog.Warningf("CopyFileRange server-side copy %s => %s outcome ambiguous: %v", copyRequest.srcPath, copyRequest.dstPath, err)
return 0, true, fuse.EIO
default:
glog.V(0).Infof("CopyFileRange server-side copy %s => %s fallback to chunk copy: %v", copyRequest.srcPath, copyRequest.dstPath, err)
return 0, false, fuse.OK
}
}
func (wfs *WFS) applyServerSideWholeFileCopyResult(fhIn, fhOut *FileHandle, dstPath util.FullPath, entry *filer_pb.Entry, sourceSize int64) {
if entry == nil {
entry = synthesizeLocalEntryForServerSideWholeFileCopy(fhIn, fhOut, sourceSize)
}
if entry == nil {
glog.Warningf("CopyFileRange server-side copy %s left no local entry to apply", dstPath)
return
}
fhOut.SetEntry(entry)
fhOut.RememberPath(dstPath)
if entry.Attributes != nil {
fhOut.contentType = entry.Attributes.Mime
}
fhOut.dirtyMetadata = false
wfs.updateServerSideWholeFileCopyMetaCache(dstPath, entry)
// Note: we intentionally skip fuseServer.InodeNotify/EntryNotify here.
// This runs inside the CopyFileRange request handler; those notifies
// would write onto the same /dev/fuse fd the kernel is still waiting
// on for this request's reply, deadlocking the mount. The kernel will
// re-read attrs once its attr-cache TTL expires.
}
func (wfs *WFS) updateServerSideWholeFileCopyMetaCache(dstPath util.FullPath, entry *filer_pb.Entry) {
if wfs.metaCache == nil || entry == nil {
return
}
dir, _ := dstPath.DirAndName()
event := metadataUpdateEvent(dir, entry)
if applyErr := wfs.applyLocalMetadataEvent(context.Background(), event); applyErr != nil {
glog.Warningf("CopyFileRange metadata update %s: %v", dstPath, applyErr)
wfs.purgeDirectoryCache(util.FullPath(dir))
}
}
func synthesizeLocalEntryForServerSideWholeFileCopy(fhIn, fhOut *FileHandle, sourceSize int64) *filer_pb.Entry {
dstEntry := fhOut.GetEntry().GetEntry()
if dstEntry == nil {
return nil
}
localEntry := proto.Clone(dstEntry).(*filer_pb.Entry)
if localEntry.Attributes == nil {
localEntry.Attributes = &filer_pb.FuseAttributes{}
}
if srcEntry := fhIn.GetEntry().GetEntry(); srcEntry != nil {
srcEntryCopy := proto.Clone(srcEntry).(*filer_pb.Entry)
localEntry.Content = srcEntryCopy.Content
localEntry.Chunks = srcEntryCopy.Chunks
if srcEntryCopy.Attributes != nil {
localEntry.Attributes.Mime = srcEntryCopy.Attributes.Mime
localEntry.Attributes.Md5 = srcEntryCopy.Attributes.Md5
}
}
copyNow := time.Now()
localEntry.Attributes.FileSize = uint64(sourceSize)
localEntry.Attributes.Mtime = copyNow.Unix()
localEntry.Attributes.MtimeNs = int32(copyNow.Nanosecond())
localEntry.Attributes.Ctime = copyNow.Unix()
localEntry.Attributes.CtimeNs = int32(copyNow.Nanosecond())
return localEntry
}
func wholeFileServerCopyCandidate(fhIn, fhOut *FileHandle, in *fuse.CopyFileRangeIn) (copyRequest wholeFileServerCopyRequest, ok bool) {
if fhIn == nil || fhOut == nil || in == nil {
glog.V(4).Infof("server-side copy: skipped (nil handle or input)")
return wholeFileServerCopyRequest{}, false
}
if fhIn.fh == fhOut.fh {
glog.V(4).Infof("server-side copy: skipped (same file handle)")
return wholeFileServerCopyRequest{}, false
}
if fhIn.dirtyMetadata || fhOut.dirtyMetadata {
glog.V(4).Infof("server-side copy: skipped (dirty metadata: in=%v out=%v)", fhIn.dirtyMetadata, fhOut.dirtyMetadata)
return wholeFileServerCopyRequest{}, false
}
if in.OffIn != 0 || in.OffOut != 0 {
glog.V(4).Infof("server-side copy: skipped (non-zero offsets: in=%d out=%d)", in.OffIn, in.OffOut)
return wholeFileServerCopyRequest{}, false
}
srcEntry := fhIn.GetEntry()
dstEntry := fhOut.GetEntry()
if srcEntry == nil || dstEntry == nil {
glog.V(4).Infof("server-side copy: skipped (nil entry: src=%v dst=%v)", srcEntry == nil, dstEntry == nil)
return wholeFileServerCopyRequest{}, false
}
if srcEntry.IsDirectory || dstEntry.IsDirectory {
glog.V(4).Infof("server-side copy: skipped (directory)")
return wholeFileServerCopyRequest{}, false
}
srcPbEntry := srcEntry.GetEntry()
dstPbEntry := dstEntry.GetEntry()
if srcPbEntry == nil || dstPbEntry == nil || srcPbEntry.Attributes == nil || dstPbEntry.Attributes == nil {
glog.V(4).Infof("server-side copy: skipped (missing entry attributes)")
return wholeFileServerCopyRequest{}, false
}
sourceSize := int64(filer.FileSize(srcPbEntry))
// go-fuse exposes CopyFileRange's return value as uint32, so the fast path
// should only claim copies that can be reported without truncation.
if sourceSize <= 0 || sourceSize > math.MaxUint32 || int64(in.Len) < sourceSize {
glog.V(4).Infof("server-side copy: skipped (size mismatch: sourceSize=%d len=%d)", sourceSize, in.Len)
return wholeFileServerCopyRequest{}, false
}
dstSize := int64(filer.FileSize(dstPbEntry))
if dstSize != 0 || len(dstPbEntry.GetChunks()) > 0 || len(dstPbEntry.Content) > 0 {
glog.V(4).Infof("server-side copy: skipped (destination not empty)")
return wholeFileServerCopyRequest{}, false
}
srcPath := fhIn.FullPath()
dstPath := fhOut.FullPath()
if srcPath == "" || dstPath == "" || srcPath == dstPath {
glog.V(4).Infof("server-side copy: skipped (invalid paths: src=%q dst=%q)", srcPath, dstPath)
return wholeFileServerCopyRequest{}, false
}
if srcPbEntry.Attributes.Inode == 0 || dstPbEntry.Attributes.Inode == 0 {
glog.V(4).Infof("server-side copy: skipped (missing inode preconditions: src=%d dst=%d)", srcPbEntry.Attributes.Inode, dstPbEntry.Attributes.Inode)
return wholeFileServerCopyRequest{}, false
}
return wholeFileServerCopyRequest{
srcPath: srcPath,
dstPath: dstPath,
sourceSize: sourceSize,
srcInode: srcPbEntry.Attributes.Inode,
srcMtime: srcPbEntry.Attributes.Mtime,
dstInode: dstPbEntry.Attributes.Inode,
dstMtime: dstPbEntry.Attributes.Mtime,
dstSize: dstSize,
sourceMime: srcPbEntry.Attributes.Mime,
sourceMd5: append([]byte(nil), srcPbEntry.Attributes.Md5...),
copyRequestID: request_id.New(),
}, true
}
func (wfs *WFS) copyEntryViaFiler(cancel <-chan struct{}, copyRequest wholeFileServerCopyRequest) (*filer_pb.Entry, serverSideWholeFileCopyOutcome, error) {
baseCtx, baseCancel := context.WithCancel(context.Background())
defer baseCancel()
if cancel != nil {
go func() {
select {
case <-cancel:
baseCancel()
case <-baseCtx.Done():
}
}()
}
postCtx, postCancel := context.WithTimeout(baseCtx, filerCopyRequestTimeout)
defer postCancel()
httpClient := util_http.GetGlobalHttpClient()
if httpClient == nil {
var err error
httpClient, err = util_http.NewGlobalHttpClient()
if err != nil {
return nil, serverSideWholeFileCopyNotCommitted, fmt.Errorf("create filer copy http client: %w", err)
}
}
copyURL := &url.URL{
Scheme: httpClient.GetHttpScheme(),
Host: wfs.getCurrentFiler().ToHttpAddress(),
Path: string(copyRequest.dstPath),
}
query := copyURL.Query()
query.Set(filer.CopyQueryParamFrom, string(copyRequest.srcPath))
query.Set(filer.CopyQueryParamOverwrite, "true")
query.Set(filer.CopyQueryParamDataOnly, "true")
query.Set(filer.CopyQueryParamRequestID, copyRequest.copyRequestID)
query.Set(filer.CopyQueryParamSourceInode, fmt.Sprintf("%d", copyRequest.srcInode))
query.Set(filer.CopyQueryParamSourceMtime, fmt.Sprintf("%d", copyRequest.srcMtime))
query.Set(filer.CopyQueryParamSourceSize, fmt.Sprintf("%d", copyRequest.sourceSize))
query.Set(filer.CopyQueryParamDestinationInode, fmt.Sprintf("%d", copyRequest.dstInode))
query.Set(filer.CopyQueryParamDestinationMtime, fmt.Sprintf("%d", copyRequest.dstMtime))
query.Set(filer.CopyQueryParamDestinationSize, fmt.Sprintf("%d", copyRequest.dstSize))
copyURL.RawQuery = query.Encode()
req, err := http.NewRequestWithContext(postCtx, http.MethodPost, copyURL.String(), nil)
if err != nil {
return nil, serverSideWholeFileCopyNotCommitted, fmt.Errorf("create filer copy request: %w", err)
}
if jwt := wfs.filerCopyJWT(); jwt != "" {
req.Header.Set("Authorization", "Bearer "+string(jwt))
}
resp, err := httpClient.Do(req)
if err != nil {
return wfs.confirmServerSideWholeFileCopyAfterAmbiguousRequest(baseCtx, copyRequest, fmt.Errorf("execute filer copy request: %w", err))
}
defer resp.Body.Close()
if resp.StatusCode != http.StatusNoContent && resp.StatusCode != http.StatusOK {
body, _ := io.ReadAll(resp.Body)
return nil, serverSideWholeFileCopyNotCommitted, fmt.Errorf("filer copy %s => %s failed: status %d: %s", copyRequest.srcPath, copyRequest.dstPath, resp.StatusCode, string(body))
}
readbackCtx, readbackCancel := context.WithTimeout(baseCtx, filerCopyReadbackTimeout)
defer readbackCancel()
entry, err := filer_pb.GetEntry(readbackCtx, wfs, copyRequest.dstPath)
if err != nil {
return nil, serverSideWholeFileCopyCommitted, fmt.Errorf("reload copied entry %s: %w", copyRequest.dstPath, err)
}
if entry == nil {
return nil, serverSideWholeFileCopyCommitted, fmt.Errorf("reload copied entry %s: not found", copyRequest.dstPath)
}
if entry.Attributes != nil && wfs.option != nil && wfs.option.UidGidMapper != nil {
entry.Attributes.Uid, entry.Attributes.Gid = wfs.option.UidGidMapper.FilerToLocal(entry.Attributes.Uid, entry.Attributes.Gid)
}
return entry, serverSideWholeFileCopyCommitted, nil
}
func (wfs *WFS) confirmServerSideWholeFileCopyAfterAmbiguousRequest(baseCtx context.Context, copyRequest wholeFileServerCopyRequest, requestErr error) (*filer_pb.Entry, serverSideWholeFileCopyOutcome, error) {
readbackCtx, readbackCancel := context.WithTimeout(baseCtx, filerCopyReadbackTimeout)
defer readbackCancel()
entry, err := filer_pb.GetEntry(readbackCtx, wfs, copyRequest.dstPath)
if err == nil && entry != nil && entryMatchesServerSideWholeFileCopy(copyRequest, entry) {
if entry.Attributes != nil && wfs.option != nil && wfs.option.UidGidMapper != nil {
entry.Attributes.Uid, entry.Attributes.Gid = wfs.option.UidGidMapper.FilerToLocal(entry.Attributes.Uid, entry.Attributes.Gid)
}
return entry, serverSideWholeFileCopyCommitted, nil
}
if err != nil {
return nil, serverSideWholeFileCopyAmbiguous, fmt.Errorf("%w; post-copy readback failed: %v", requestErr, err)
}
if entry == nil {
return nil, serverSideWholeFileCopyAmbiguous, fmt.Errorf("%w; destination %s was not readable after the ambiguous request", requestErr, copyRequest.dstPath)
}
return nil, serverSideWholeFileCopyAmbiguous, fmt.Errorf("%w; destination %s did not match the requested copy after the ambiguous request", requestErr, copyRequest.dstPath)
}
func entryMatchesServerSideWholeFileCopy(copyRequest wholeFileServerCopyRequest, entry *filer_pb.Entry) bool {
if entry == nil || entry.Attributes == nil {
return false
}
if copyRequest.dstInode != 0 && entry.Attributes.Inode != copyRequest.dstInode {
return false
}
if entry.Attributes.FileSize != uint64(copyRequest.sourceSize) {
return false
}
if copyRequest.sourceMime != "" && entry.Attributes.Mime != copyRequest.sourceMime {
return false
}
if len(copyRequest.sourceMd5) > 0 && !bytes.Equal(entry.Attributes.Md5, copyRequest.sourceMd5) {
return false
}
return true
}
func (wfs *WFS) filerCopyJWT() security.EncodedJwt {
if wfs.option == nil || len(wfs.option.FilerSigningKey) == 0 {
return ""
}
return security.GenJwtForFilerServer(wfs.option.FilerSigningKey, wfs.option.FilerSigningExpiresAfterSec)
}