X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/d1327f9456b561c059c18c76f63391728ce1780d..d137cbc6cfdcc541216e68d414c535626e4d8916:/services/keepstore/handlers.go diff --git a/services/keepstore/handlers.go b/services/keepstore/handlers.go index 2b96dbc582..f698982415 100644 --- a/services/keepstore/handlers.go +++ b/services/keepstore/handlers.go @@ -20,6 +20,7 @@ import ( "regexp" "runtime" "strconv" + "strings" "sync" "time" ) @@ -53,6 +54,9 @@ func MakeRESTRouter() *mux.Router { // Replace the current trash queue. rest.HandleFunc(`/trash`, TrashHandler).Methods("PUT") + // Untrash moves blocks from trash back into store + rest.HandleFunc(`/untrash/{hash:[0-9a-f]{32}}`, UntrashHandler).Methods("PUT") + // Any request which does not match any of these routes gets // 400 Bad Request. rest.NotFoundHandler = http.HandlerFunc(BadRequestHandler) @@ -75,18 +79,61 @@ func GetBlockHandler(resp http.ResponseWriter, req *http.Request) { } } - block, err := GetBlock(mux.Vars(req)["hash"]) + // TODO: Probe volumes to check whether the block _might_ + // exist. Some volumes/types could support a quick existence + // check without causing other operations to suffer. If all + // volumes support that, and assure us the block definitely + // isn't here, we can return 404 now instead of waiting for a + // buffer. + + buf, err := getBufferForResponseWriter(resp, bufs, BlockSize) if err != nil { - // This type assertion is safe because the only errors - // GetBlock can return are DiskHashError or NotFoundError. - http.Error(resp, err.Error(), err.(*KeepError).HTTPCode) + http.Error(resp, err.Error(), http.StatusServiceUnavailable) return } - defer bufs.Put(block) + defer bufs.Put(buf) - resp.Header().Set("Content-Length", strconv.Itoa(len(block))) + size, err := GetBlock(mux.Vars(req)["hash"], buf, resp) + if err != nil { + code := http.StatusInternalServerError + if err, ok := err.(*KeepError); ok { + code = err.HTTPCode + } + http.Error(resp, err.Error(), code) + return + } + + resp.Header().Set("Content-Length", strconv.Itoa(size)) resp.Header().Set("Content-Type", "application/octet-stream") - resp.Write(block) + resp.Write(buf[:size]) +} + +// Get a buffer from the pool -- but give up and return a non-nil +// error if resp implements http.CloseNotifier and tells us that the +// client has disconnected before we get a buffer. +func getBufferForResponseWriter(resp http.ResponseWriter, bufs *bufferPool, bufSize int) ([]byte, error) { + var closeNotifier <-chan bool + if resp, ok := resp.(http.CloseNotifier); ok { + closeNotifier = resp.CloseNotify() + } + var buf []byte + bufReady := make(chan []byte) + go func() { + bufReady <- bufs.Get(bufSize) + close(bufReady) + }() + select { + case buf = <-bufReady: + return buf, nil + case <-closeNotifier: + go func() { + // Even if closeNotifier happened first, we + // need to keep waiting for our buf so we can + // return it to the pool. + bufs.Put(<-bufReady) + }() + return nil, ErrClientDisconnect + } } // PutBlockHandler is a HandleFunc to address Put block requests. @@ -112,15 +159,20 @@ func PutBlockHandler(resp http.ResponseWriter, req *http.Request) { return } - buf := bufs.Get(int(req.ContentLength)) - _, err := io.ReadFull(req.Body, buf) + buf, err := getBufferForResponseWriter(resp, bufs, int(req.ContentLength)) + if err != nil { + http.Error(resp, err.Error(), http.StatusServiceUnavailable) + return + } + + _, err = io.ReadFull(req.Body, buf) if err != nil { http.Error(resp, err.Error(), 500) bufs.Put(buf) return } - err = PutBlock(buf, hash) + replication, err := PutBlock(buf, hash) bufs.Put(buf) if err != nil { @@ -137,6 +189,7 @@ func PutBlockHandler(resp http.ResponseWriter, req *http.Request) { expiry := time.Now().Add(blobSignatureTTL) returnHash = SignLocator(returnHash, apiToken, expiry) } + resp.Header().Set("X-Keep-Replicas-Stored", strconv.Itoa(replication)) resp.Write([]byte(returnHash + "\n")) } @@ -294,7 +347,7 @@ func DeleteHandler(resp http.ResponseWriter, req *http.Request) { Failed int `json:"copies_failed"` } for _, vol := range KeepVM.AllWritable() { - if err := vol.Delete(hash); err == nil { + if err := vol.Trash(hash); err == nil { result.Deleted++ } else if os.IsNotExist(err) { continue @@ -429,7 +482,53 @@ func TrashHandler(resp http.ResponseWriter, req *http.Request) { trashq.ReplaceQueue(tlist) } -// ============================== +// UntrashHandler processes "PUT /untrash/{hash:[0-9a-f]{32}}" requests for the data manager. +func UntrashHandler(resp http.ResponseWriter, req *http.Request) { + // Reject unauthorized requests. + if !IsDataManagerToken(GetApiToken(req)) { + http.Error(resp, UnauthorizedError.Error(), UnauthorizedError.HTTPCode) + return + } + + hash := mux.Vars(req)["hash"] + + if len(KeepVM.AllWritable()) == 0 { + http.Error(resp, "No writable volumes", http.StatusNotFound) + return + } + + var untrashedOn, failedOn []string + var numNotFound int + for _, vol := range KeepVM.AllWritable() { + err := vol.Untrash(hash) + + if os.IsNotExist(err) { + numNotFound++ + } else if err != nil { + log.Printf("Error untrashing %v on volume %v", hash, vol.String()) + failedOn = append(failedOn, vol.String()) + } else { + log.Printf("Untrashed %v on volume %v", hash, vol.String()) + untrashedOn = append(untrashedOn, vol.String()) + } + } + + if numNotFound == len(KeepVM.AllWritable()) { + http.Error(resp, "Block not found on any of the writable volumes", http.StatusNotFound) + return + } + + if len(failedOn) == len(KeepVM.AllWritable()) { + http.Error(resp, "Failed to untrash on all writable volumes", http.StatusInternalServerError) + } else { + respBody := "Successfully untrashed on: " + strings.Join(untrashedOn, ",") + if len(failedOn) > 0 { + respBody += "; Failed to untrash on: " + strings.Join(failedOn, ",") + } + resp.Write([]byte(respBody)) + } +} + // GetBlock and PutBlock implement lower-level code for handling // blocks by rooting through volumes connected to the local machine. // Once the handler has determined that system policy permits the @@ -440,24 +539,21 @@ func TrashHandler(resp http.ResponseWriter, req *http.Request) { // should be the only part of the code that cares about which volume a // block is stored on, so it should be responsible for figuring out // which volume to check for fetching blocks, storing blocks, etc. -// ============================== -// GetBlock fetches and returns the block identified by "hash". -// -// On success, GetBlock returns a byte slice with the block data, and -// a nil error. +// GetBlock fetches the block identified by "hash" into the provided +// buf, and returns the data size. // // If the block cannot be found on any volume, returns NotFoundError. // // If the block found does not have the correct MD5 hash, returns // DiskHashError. // -func GetBlock(hash string) ([]byte, error) { +func GetBlock(hash string, buf []byte, resp http.ResponseWriter) (int, error) { // Attempt to read the requested hash from a keep volume. errorToCaller := NotFoundError for _, vol := range KeepVM.AllReadable() { - buf, err := vol.Get(hash) + size, err := vol.Get(hash, buf) if err != nil { // IsNotExist is an expected error and may be // ignored. All other errors are logged. In @@ -471,23 +567,22 @@ func GetBlock(hash string) ([]byte, error) { } // Check the file checksum. // - filehash := fmt.Sprintf("%x", md5.Sum(buf)) + filehash := fmt.Sprintf("%x", md5.Sum(buf[:size])) if filehash != hash { // TODO: Try harder to tell a sysadmin about // this. log.Printf("%s: checksum mismatch for request %s (actual %s)", vol, hash, filehash) errorToCaller = DiskHashError - bufs.Put(buf) continue } if errorToCaller == DiskHashError { log.Printf("%s: checksum mismatch for request %s but a good copy was found on another volume and returned", vol, hash) } - return buf, nil + return size, nil } - return nil, errorToCaller + return 0, errorToCaller } // PutBlock Stores the BLOCK (identified by the content id HASH) in Keep. @@ -517,40 +612,40 @@ func GetBlock(hash string) ([]byte, error) { // all writes failed). The text of the error message should // provide as much detail as possible. // -func PutBlock(block []byte, hash string) error { +func PutBlock(block []byte, hash string) (int, error) { // Check that BLOCK's checksum matches HASH. blockhash := fmt.Sprintf("%x", md5.Sum(block)) if blockhash != hash { log.Printf("%s: MD5 checksum %s did not match request", hash, blockhash) - return RequestHashError + return 0, RequestHashError } // If we already have this data, it's intact on disk, and we // can update its timestamp, return success. If we have // different data with the same hash, return failure. - if err := CompareAndTouch(hash, block); err == nil || err == CollisionError { - return err + if n, err := CompareAndTouch(hash, block); err == nil || err == CollisionError { + return n, err } // Choose a Keep volume to write to. // If this volume fails, try all of the volumes in order. if vol := KeepVM.NextWritable(); vol != nil { if err := vol.Put(hash, block); err == nil { - return nil // success! + return vol.Replication(), nil // success! } } writables := KeepVM.AllWritable() if len(writables) == 0 { log.Print("No writable volumes.") - return FullError + return 0, FullError } allFull := true for _, vol := range writables { err := vol.Put(hash, block) if err == nil { - return nil // success! + return vol.Replication(), nil // success! } if err != FullError { // The volume is not full but the @@ -563,17 +658,18 @@ func PutBlock(block []byte, hash string) error { if allFull { log.Print("All volumes are full.") - return FullError + return 0, FullError } // Already logged the non-full errors. - return GenericError + return 0, GenericError } -// CompareAndTouch returns nil if one of the volumes already has the -// given content and it successfully updates the relevant block's -// modification time in order to protect it from premature garbage -// collection. -func CompareAndTouch(hash string, buf []byte) error { +// CompareAndTouch returns the current replication level if one of the +// volumes already has the given content and it successfully updates +// the relevant block's modification time in order to protect it from +// premature garbage collection. Otherwise, it returns a non-nil +// error. +func CompareAndTouch(hash string, buf []byte) (int, error) { var bestErr error = NotFoundError for _, vol := range KeepVM.AllWritable() { if err := vol.Compare(hash, buf); err == CollisionError { @@ -583,7 +679,7 @@ func CompareAndTouch(hash string, buf []byte) error { // both, so there's no point writing it even // on a different volume.) log.Printf("%s: Compare(%s): %s", vol, hash, err) - return err + return 0, err } else if os.IsNotExist(err) { // Block does not exist. This is the only // "normal" error: we don't log anything. @@ -601,9 +697,9 @@ func CompareAndTouch(hash string, buf []byte) error { continue } // Compare and Touch both worked --> done. - return nil + return vol.Replication(), nil } - return bestErr + return 0, bestErr } var validLocatorRe = regexp.MustCompile(`^[0-9a-f]{32}$`)