Files
seaweedfs/weed/s3api/s3api_object_handlers_copy_part_sse.go
T
Chris Lu 7bb0a1c127 s3: replay a delete whose reply the transport dropped (#11022)
* s3: stop retrying a delete the filer refused for a non-empty folder

The filer looked and the children are there, so the answer will not change.
retryFilerOp spent six attempts and up to 3.1s of backoff on it before the
caller could act on the condition it was already holding.

Claude-Session: https://claude.ai/code/session_01XqaJrwgXQ5GSUpyzRbe5nD

* s3: thread the request context through the unversioned delete path

doDeleteEntry issued every DeleteEntry on context.Background(), so an S3
client that hung up left the gateway working on its behalf, out of reach of
both cancellation and the per-request retry allowance that
DeleteMultipleObjectsHandler installs.

Claude-Session: https://claude.ai/code/session_01XqaJrwgXQ5GSUpyzRbe5nD

* s3: treat a cancelled filer RPC as terminal, not transient

isRetryableFilerErr matched context.Canceled and DeadlineExceeded by
sentinel, which only holds while the error is still local. Once it has
crossed gRPC it is a status, so an abandoned request was retried six times
on behalf of a caller that had already gone.

Claude-Session: https://claude.ai/code/session_01XqaJrwgXQ5GSUpyzRbe5nD

* s3: replay a delete whose reply the transport dropped

A delete is idempotent at the filer, which answers an entry that is already
gone with an empty resp.Error, so a reply lost in transit can be reissued
rather than surfaced. Surfaced, it becomes a 500 on the bucket delete, which
boto3 resends and is then answered NoSuchBucket, or a per-key InternalError
inside the 200 of a multi-object delete, which no SDK retries at all.

The replay runs through retryFilerOp, so it draws on the allowance the
request already installs rather than paying a backoff per key, and stops for
a caller that has gone. rm and rmObject re-enter WithFilerClient per attempt,
so each one walks the failover list again on a connection the failed attempt
had invalidated; the multi-object loop holds one client for the batch, so
there the replay reuses it.

Classification stays structural. The filer reports its own refusals in
resp.Error, which carries no status and has the deleted path - and, for a
recursive delete, the children it stopped on - formatted into it, so no key
name can steer the decision either way.

rm and rmObject now take the caller's context. Cleanup and rollback paths
pass context.Background() deliberately: they have to run whether or not the
caller is still waiting.

Claude-Session: https://claude.ai/code/session_01XqaJrwgXQ5GSUpyzRbe5nD

* s3: share one retry allowance across multipart completion cleanup

The unused-entry loop deletes once per entry, and each delete now retries,
so a filer that stays unavailable held the response for 3.1s per entry after
the object was already committed.

Claude-Session: https://claude.ai/code/session_01XqaJrwgXQ5GSUpyzRbe5nD
2026-08-28 14:30:21 -07:00

433 lines
16 KiB
Go

package s3api
import (
"context"
"encoding/base64"
"encoding/json"
"errors"
"fmt"
"io"
"net/http"
"sort"
"strconv"
"google.golang.org/grpc/codes"
"google.golang.org/grpc/status"
"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/s3api/s3_constants"
"github.com/seaweedfs/seaweedfs/weed/s3api/s3err"
)
// errCopySourceSSEUnsupported is returned by openSourcePlaintextReader when
// the source object's SSE type is not yet implemented in the UploadPartCopy
// slow path. Callers map it to a 501 NotImplemented S3 response so clients
// can distinguish "we will not handle this shape" from "the server failed".
var errCopySourceSSEUnsupported = errors.New("UploadPartCopy source SSE type not yet supported")
// isTransientFilerError reports whether an error talking to the filer is
// retryable from the client's perspective (filer briefly unreachable, leader
// election in flight, deadline exceeded, etc.). Such errors should map to a
// 503 ServiceUnavailable response so SDK retry logic engages, rather than a
// 500 InternalError which most clients treat as fatal.
func isTransientFilerError(err error) bool {
if err == nil {
return false
}
if errors.Is(err, context.DeadlineExceeded) || errors.Is(err, context.Canceled) {
return true
}
if s, ok := status.FromError(err); ok {
switch s.Code() {
case codes.Unavailable, codes.DeadlineExceeded, codes.ResourceExhausted, codes.Aborted:
return true
}
}
return false
}
// uploadEntryHasSSE reports whether the multipart upload entry was created
// with any server-side encryption configured (SSE-S3 or SSE-KMS — explicit at
// CreateMultipartUpload time or applied as bucket default). It is used to
// decide whether UploadPartCopy must re-encrypt source bytes for the
// destination, rather than copying them as raw bytes (the fast path).
func uploadEntryHasSSE(uploadEntry *filer_pb.Entry) bool {
if uploadEntry == nil || uploadEntry.Extended == nil {
return false
}
if _, ok := uploadEntry.Extended[s3_constants.SeaweedFSSSEKMSKeyID]; ok {
return true
}
if v, ok := uploadEntry.Extended[s3_constants.SeaweedFSSSES3Encryption]; ok && string(v) == s3_constants.SSEAlgorithmAES256 {
return true
}
return false
}
func uploadEntryHasChecksum(uploadEntry *filer_pb.Entry) bool {
if uploadEntry == nil || uploadEntry.Extended == nil {
return false
}
headerName := string(uploadEntry.Extended[s3_constants.ExtChecksumAlgorithm])
return checksumAlgorithmFromHeaderName(headerName) != ChecksumAlgorithmNone
}
// sourceEntryIsEncrypted reports whether the source object's chunks are
// ciphertext on disk and therefore cannot be raw-copied — they must be
// decrypted on read. That is SSE, and also -encryptVolumeData, whose per-chunk
// key only ever decrypts a whole chunk and whose chunks carry no ETag for the
// copied part to report.
func sourceEntryIsEncrypted(srcEntry *filer_pb.Entry) bool {
if srcEntry == nil {
return false
}
for _, c := range srcEntry.GetChunks() {
if c.GetSseType() != filer_pb.SSEType_NONE || len(c.GetCipherKey()) > 0 {
return true
}
}
if srcEntry.Extended != nil {
if _, ok := srcEntry.Extended[s3_constants.SeaweedFSSSES3Key]; ok {
return true
}
if _, ok := srcEntry.Extended[s3_constants.SeaweedFSSSEKMSKey]; ok {
return true
}
if _, ok := srcEntry.Extended[s3_constants.AmzServerSideEncryptionCustomerAlgorithm]; ok {
return true
}
}
return false
}
// readCloserAdapter pairs an arbitrary io.Reader with an io.Closer so callers
// can release the original underlying source even when the inner Reader (e.g.
// cipher.StreamReader, io.LimitReader) does not implement io.Closer.
type readCloserAdapter struct {
io.Reader
closer io.Closer
}
func (r *readCloserAdapter) Close() error {
if r.closer == nil {
return nil
}
return r.closer.Close()
}
// openSourcePlaintextReader returns a reader yielding the source object's
// plaintext bytes for [startOffset, endOffset], applying any necessary SSE
// decryption based on the source entry's metadata.
//
// Used by CopyObjectPartHandler when source or destination is SSE-encrypted:
// the fast raw-chunk-copy path leaves destination chunks SseType=NONE and
// completedMultipartChunk's NONE→SSE_S3 backfill (PR #9224) then writes
// destination-baseIV-derived metadata onto bytes that were actually encrypted
// with the source's key — producing deterministic byte corruption on GET (#8908).
//
// Returns errCopySourceSSEUnsupported when the source's SSE type is not yet
// implemented in this slow path (SSE-KMS, SSE-C). Callers should map that
// sentinel to a 501 NotImplemented S3 response rather than collapsing it to
// 500 InternalError, so clients can distinguish "we will not handle this
// shape" from "the server failed".
func (s3a *S3ApiServer) openSourcePlaintextReader(
ctx context.Context,
srcEntry *filer_pb.Entry,
startOffset, endOffset int64,
) (io.ReadCloser, error) {
if srcEntry == nil {
return nil, fmt.Errorf("nil source entry")
}
if endOffset < startOffset {
return io.NopCloser(io.LimitReader(emptyReader{}, 0)), nil
}
sliceLen := endOffset - startOffset + 1
switch s3a.detectPrimarySSEType(srcEntry) {
case s3_constants.SSETypeS3:
return s3a.openSSES3SourcePlaintextReader(ctx, srcEntry, startOffset, sliceLen)
case s3_constants.SSETypeKMS:
return nil, fmt.Errorf("%w: UploadPartCopy from SSE-KMS source", errCopySourceSSEUnsupported)
case s3_constants.SSETypeC:
return nil, fmt.Errorf("%w: UploadPartCopy from SSE-C source", errCopySourceSSEUnsupported)
default:
// Plaintext or volume-encrypted source: the chunk stream seeks, so ask
// it for the range rather than reading and discarding the prefix.
raw, err := s3a.getEncryptedStreamFromVolumes(ctx, srcEntry, startOffset, sliceLen)
if err != nil {
return nil, fmt.Errorf("open source: %w", err)
}
return raw, nil
}
}
// openSSES3SourcePlaintextReader builds a decrypted reader for an SSE-S3
// source. It reuses buildMultipartSSES3Reader, which decrypts each chunk
// independently using its per-chunk metadata — correct for both multipart-SSE
// objects (multiple SSE-S3 chunks) and single-part SSE-S3 objects whose single
// chunk also carries per-chunk metadata after PR #9211.
//
// For older single-part SSE-S3 objects whose chunks lack per-chunk metadata,
// this falls back to the entry-level SSE-S3 key + the entry's stored IV,
// matching the read path's single-part fallback.
func (s3a *S3ApiServer) openSSES3SourcePlaintextReader(
ctx context.Context,
srcEntry *filer_pb.Entry,
startOffset, sliceLen int64,
) (io.ReadCloser, error) {
chunks := srcEntry.GetChunks()
hasPerChunkSSE := false
for _, c := range chunks {
if c.GetSseType() == filer_pb.SSEType_SSE_S3 && len(c.GetSseMetadata()) > 0 {
hasPerChunkSSE = true
break
}
}
if hasPerChunkSSE {
sortedChunks := make([]*filer_pb.FileChunk, len(chunks))
copy(sortedChunks, chunks)
sort.Slice(sortedChunks, func(i, j int) bool {
return sortedChunks[i].GetOffset() < sortedChunks[j].GetOffset()
})
decReader, err := buildMultipartSSES3Reader(
sortedChunks,
GetSSES3KeyManager(),
func(c *filer_pb.FileChunk) (io.ReadCloser, error) {
return s3a.createEncryptedChunkReader(ctx, c)
},
)
if err != nil {
return nil, fmt.Errorf("build SSE-S3 source reader: %w", err)
}
// buildMultipartSSES3Reader returns a *lazyMultipartChunkReader whose
// Close() releases the live chunk body. Use it as the closer.
var closer io.Closer
if rc, ok := decReader.(io.Closer); ok {
closer = rc
}
return applyRange(&readCloserAdapter{Reader: decReader, closer: closer}, startOffset, sliceLen)
}
// Legacy single-part fallback: entry-level SeaweedFSSSES3Key + entry IV.
keyData, ok := srcEntry.Extended[s3_constants.SeaweedFSSSES3Key]
if !ok || len(keyData) == 0 {
return nil, fmt.Errorf("SSE-S3 source has no per-chunk metadata and no entry-level SSE-S3 key")
}
keyManager := GetSSES3KeyManager()
sseS3Key, err := DeserializeSSES3Metadata(keyData, keyManager)
if err != nil {
return nil, fmt.Errorf("deserialize entry-level SSE-S3 key: %w", err)
}
iv, err := GetSSES3IV(srcEntry, sseS3Key, keyManager)
if err != nil {
return nil, fmt.Errorf("get SSE-S3 IV: %w", err)
}
encStream, err := s3a.getEncryptedStreamFromVolumes(ctx, srcEntry, 0, int64(filer.FileSize(srcEntry)))
if err != nil {
return nil, fmt.Errorf("open ciphertext source: %w", err)
}
dec, err := CreateSSES3DecryptedReader(encStream, sseS3Key, iv)
if err != nil {
encStream.Close()
return nil, fmt.Errorf("create SSE-S3 decrypted reader: %w", err)
}
rc, ok := dec.(io.ReadCloser)
if !ok {
rc = &readCloserAdapter{Reader: dec, closer: encStream}
}
return applyRange(rc, startOffset, sliceLen)
}
// applyRange skips startOffset bytes from src and limits the result to
// sliceLen bytes. The returned ReadCloser closes the underlying source.
func applyRange(src io.ReadCloser, startOffset, sliceLen int64) (io.ReadCloser, error) {
if startOffset > 0 {
if _, err := io.CopyN(io.Discard, src, startOffset); err != nil {
src.Close()
return nil, fmt.Errorf("skip to range start %d: %w", startOffset, err)
}
}
if sliceLen <= 0 {
return &readCloserAdapter{Reader: io.LimitReader(src, 0), closer: src}, nil
}
return &readCloserAdapter{Reader: io.LimitReader(src, sliceLen), closer: src}, nil
}
// emptyReader yields no bytes. Used for empty-range UploadPartCopy.
type emptyReader struct{}
func (emptyReader) Read([]byte) (int, error) { return 0, io.EOF }
// applyDestSSEHeadersToCopyRequest stages the destination's SSE setup on the
// (cloned) request so that putToFiler's existing handleAllSSEEncryption picks
// it up. The upload-entry markers (laid down at CreateMultipartUpload) bind
// every part of the upload to the same key+baseIV, matching PutObjectPart.
func (s3a *S3ApiServer) applyDestSSEHeadersToCopyRequest(
r *http.Request, uploadEntry *filer_pb.Entry, uploadID string,
) error {
if uploadEntry == nil || uploadEntry.Extended == nil {
return nil
}
if keyIDBytes, hasKMS := uploadEntry.Extended[s3_constants.SeaweedFSSSEKMSKeyID]; hasKMS {
// Mirror the SSE-KMS branch of PutObjectPartHandler: stage
// X-Amz-Server-Side-Encryption=aws:kms plus the key ID, encryption
// context, bucket-key flag and base IV onto the request.
keyID := string(keyIDBytes)
bucketKeyEnabled := false
if v, ok := uploadEntry.Extended[s3_constants.SeaweedFSSSEKMSBucketKeyEnabled]; ok && string(v) == "true" {
bucketKeyEnabled = true
}
var encryptionContext map[string]string
if cb, ok := uploadEntry.Extended[s3_constants.SeaweedFSSSEKMSEncryptionContext]; ok {
if err := json.Unmarshal(cb, &encryptionContext); err != nil {
glog.Errorf("UploadPartCopy: failed to parse SSE-KMS context for upload %s: %v", uploadID, err)
encryptionContext = nil
}
}
if len(encryptionContext) == 0 {
// Bucket and object are populated on the cloned request; reuse
// the same builder PutObjectPartHandler does.
bucket, object := s3_constants.GetBucketAndObject(r)
encryptionContext = BuildEncryptionContext(bucket, object, bucketKeyEnabled)
}
var baseIV []byte
if ivBytes, ok := uploadEntry.Extended[s3_constants.SeaweedFSSSEKMSBaseIV]; ok {
decoded, decErr := base64.StdEncoding.DecodeString(string(ivBytes))
if decErr != nil || len(decoded) != s3_constants.AESBlockSize {
return fmt.Errorf("invalid SSE-KMS base IV on upload %s", uploadID)
}
baseIV = decoded
} else {
return fmt.Errorf("no SSE-KMS base IV on upload %s", uploadID)
}
r.Header.Set(s3_constants.AmzServerSideEncryption, "aws:kms")
r.Header.Set(s3_constants.AmzServerSideEncryptionAwsKmsKeyId, keyID)
if bucketKeyEnabled {
r.Header.Set(s3_constants.AmzServerSideEncryptionBucketKeyEnabled, "true")
}
if len(encryptionContext) > 0 {
if cj, err := json.Marshal(encryptionContext); err == nil {
r.Header.Set(s3_constants.AmzServerSideEncryptionContext, base64.StdEncoding.EncodeToString(cj))
}
}
r.Header.Set(s3_constants.SeaweedFSSSEKMSBaseIVHeader, base64.StdEncoding.EncodeToString(baseIV))
return nil
}
// SSE-S3 path: reuse the existing PutObjectPart helper unchanged. It is
// pure header manipulation on r and does not touch S3ApiServer state.
return s3a.handleSSES3MultipartHeaders(r, uploadEntry, uploadID)
}
func applyDestChecksumHeaderToCopyRequest(r *http.Request, uploadEntry *filer_pb.Entry) {
if uploadEntry == nil || uploadEntry.Extended == nil {
return
}
headerName := string(uploadEntry.Extended[s3_constants.ExtChecksumAlgorithm])
if algorithm := checksumAlgorithmNameFromHeaderName(headerName); algorithm != "" {
// Drop any inherited sdk-checksum selector; it outranks the header we set.
r.Header.Del(s3_constants.AmzSdkChecksumAlgorithm)
r.Header.Set(s3_constants.AmzChecksumAlgorithm, algorithm)
}
}
func fakeContentRequest(orig *http.Request, body io.ReadCloser, contentLength int64) *http.Request {
cloned := orig.Clone(orig.Context())
cloned.Body = body
cloned.ContentLength = contentLength
if cloned.Header == nil {
cloned.Header = http.Header{}
} else {
cloned.Header = cloned.Header.Clone()
}
cloned.Header.Set("Content-Length", strconv.FormatInt(contentLength, 10))
cloned.Header.Del("X-Amz-Copy-Source")
cloned.Header.Del("X-Amz-Copy-Source-Range")
cloned.Header.Del("X-Amz-Metadata-Directive")
cloned.Header.Del("X-Amz-Tagging-Directive")
// Content-Md5 cannot be reproduced from the source plaintext without
// streaming it once first; clear it so putToFiler doesn't validate.
cloned.Header.Del("Content-Md5")
return cloned
}
// copyObjectPartViaReencryption is the UploadPartCopy slow path: it re-streams the
// source range through putToFiler so the destination's SSE re-encryption and/or
// requested checksum are produced on write. A raw chunk copy can't: it would leave
// dest chunks under the source key (corrupt GET) and parts with no checksum
// (completion fails).
func (s3a *S3ApiServer) copyObjectPartViaReencryption(
r *http.Request,
srcEntry *filer_pb.Entry,
startOffset, endOffset int64,
dstBucket, dstObject, uploadID string,
partID int,
uploadEntry *filer_pb.Entry,
) (etag string, sseMetadata SSEResponseMetadata, errCode s3err.ErrorCode) {
if endOffset < startOffset && !uploadEntryHasChecksum(uploadEntry) {
tag, code := s3a.writeEmptyCopyPart(dstBucket, uploadID, partID)
return tag, SSEResponseMetadata{}, code
}
sliceLen := int64(0)
if endOffset >= startOffset {
sliceLen = endOffset - startOffset + 1
}
srcReader, err := s3a.openSourcePlaintextReader(r.Context(), srcEntry, startOffset, endOffset)
if err != nil {
glog.Errorf("UploadPartCopy: open source plaintext reader: %v", err)
// Distinguish "we will not handle this shape" (501) from "the server
// failed" (500). SSE-KMS / SSE-C source support in this slow path is
// staged work; the explicit error lets clients see it as a feature
// gap rather than a server fault.
if errors.Is(err, errCopySourceSSEUnsupported) {
return "", SSEResponseMetadata{}, s3err.ErrNotImplemented
}
return "", SSEResponseMetadata{}, s3err.ErrInternalError
}
defer srcReader.Close()
cloned := fakeContentRequest(r, srcReader, sliceLen)
if err := s3a.applyDestSSEHeadersToCopyRequest(cloned, uploadEntry, uploadID); err != nil {
glog.Errorf("UploadPartCopy: apply destination SSE headers: %v", err)
return "", SSEResponseMetadata{}, s3err.ErrInternalError
}
applyDestChecksumHeaderToCopyRequest(cloned, uploadEntry)
filePath := s3a.genPartUploadPath(dstBucket, uploadID, partID)
tag, code, putSSE := s3a.putToFiler(cloned, filePath, srcReader, dstBucket, "", partID, 0, nil, false, s3a.toFilerPath(dstBucket, dstObject))
if code != s3err.ErrNone {
return "", SSEResponseMetadata{}, code
}
return tag, putSSE, s3err.ErrNone
}
// writeEmptyCopyPart writes a 0-byte part entry for an empty UploadPartCopy
// range, mirroring the legacy fast path's handling of endOffset < startOffset.
func (s3a *S3ApiServer) writeEmptyCopyPart(dstBucket, uploadID string, partID int) (string, s3err.ErrorCode) {
uploadDir, partName := s3a.copyPartLocation(dstBucket, uploadID, partID)
if exists, _ := s3a.exists(uploadDir, partName, false); exists {
if err := s3a.rm(context.Background(), uploadDir, partName, false, false); err != nil {
return "", s3err.ErrInternalError
}
}
if err := s3a.mkFile(uploadDir, partName, nil, func(e *filer_pb.Entry) {
if e.Attributes == nil {
e.Attributes = &filer_pb.FuseAttributes{}
}
e.Attributes.FileSize = 0
}); err != nil {
return "", s3err.ErrInternalError
}
const emptyMD5Hex = "d41d8cd98f00b204e9800998ecf8427e"
return emptyMD5Hex, s3err.ErrNone
}