X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/5fa837af6e73abbcf70e66c6de785e3a24259328..d137cbc6cfdcc541216e68d414c535626e4d8916:/services/keepstore/handlers.go diff --git a/services/keepstore/handlers.go b/services/keepstore/handlers.go index d1169d5df9..f698982415 100644 --- a/services/keepstore/handlers.go +++ b/services/keepstore/handlers.go @@ -8,7 +8,6 @@ package main // StatusHandler (GET /status.json) import ( - "bytes" "container/list" "crypto/md5" "encoding/json" @@ -21,6 +20,7 @@ import ( "regexp" "runtime" "strconv" + "strings" "sync" "time" ) @@ -54,6 +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) @@ -61,12 +64,14 @@ func MakeRESTRouter() *mux.Router { return rest } +// BadRequestHandler is a HandleFunc to address bad requests. func BadRequestHandler(w http.ResponseWriter, r *http.Request) { http.Error(w, BadRequestError.Error(), BadRequestError.HTTPCode) } +// GetBlockHandler is a HandleFunc to address Get block requests. func GetBlockHandler(resp http.ResponseWriter, req *http.Request) { - if enforce_permissions { + if enforcePermissions { locator := req.URL.Path[1:] // strip leading slash if err := VerifySignature(locator, GetApiToken(req)); err != nil { http.Error(resp, err.Error(), err.(*KeepError).HTTPCode) @@ -74,20 +79,64 @@ func GetBlockHandler(resp http.ResponseWriter, req *http.Request) { } } - block, err := GetBlock(mux.Vars(req)["hash"], false) + // 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. func PutBlockHandler(resp http.ResponseWriter, req *http.Request) { hash := mux.Vars(req)["hash"] @@ -100,7 +149,7 @@ func PutBlockHandler(resp http.ResponseWriter, req *http.Request) { return } - if req.ContentLength > BLOCKSIZE { + if req.ContentLength > BlockSize { http.Error(resp, TooLongError.Error(), TooLongError.HTTPCode) return } @@ -110,15 +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 { @@ -129,18 +183,17 @@ func PutBlockHandler(resp http.ResponseWriter, req *http.Request) { // Success; add a size hint, sign the locator if possible, and // return it to the client. - return_hash := fmt.Sprintf("%s+%d", hash, req.ContentLength) - api_token := GetApiToken(req) - if PermissionSecret != nil && api_token != "" { - expiry := time.Now().Add(blob_signature_ttl) - return_hash = SignLocator(return_hash, api_token, expiry) - } - resp.Write([]byte(return_hash + "\n")) + returnHash := fmt.Sprintf("%s+%d", hash, req.ContentLength) + apiToken := GetApiToken(req) + if PermissionSecret != nil && apiToken != "" { + 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")) } -// IndexHandler -// A HandleFunc to address /index and /index/{prefix} requests. -// +// IndexHandler is a HandleFunc to address /index and /index/{prefix} requests. func IndexHandler(resp http.ResponseWriter, req *http.Request) { // Reject unauthorized requests. if !IsDataManagerToken(GetApiToken(req)) { @@ -178,46 +231,43 @@ func IndexHandler(resp http.ResponseWriter, req *http.Request) { // * device_num (an integer identifying the underlying filesystem) // * bytes_free // * bytes_used -// -type VolumeStatus struct { - MountPoint string `json:"mount_point"` - DeviceNum uint64 `json:"device_num"` - BytesFree uint64 `json:"bytes_free"` - BytesUsed uint64 `json:"bytes_used"` -} +// PoolStatus struct type PoolStatus struct { Alloc uint64 `json:"BytesAllocated"` Cap int `json:"BuffersMax"` Len int `json:"BuffersInUse"` } +// NodeStatus struct type NodeStatus struct { - Volumes []*VolumeStatus `json:"volumes"` + Volumes []*VolumeStatus `json:"volumes"` BufferPool PoolStatus + PullQueue WorkQueueStatus + TrashQueue WorkQueueStatus Memory runtime.MemStats } var st NodeStatus var stLock sync.Mutex + +// StatusHandler addresses /status.json requests. func StatusHandler(resp http.ResponseWriter, req *http.Request) { stLock.Lock() - ReadNodeStatus(&st) + readNodeStatus(&st) jstat, err := json.Marshal(&st) stLock.Unlock() if err == nil { resp.Write(jstat) } else { - log.Printf("json.Marshal: %s\n", err) - log.Printf("NodeStatus = %v\n", &st) + log.Printf("json.Marshal: %s", err) + log.Printf("NodeStatus = %v", &st) http.Error(resp, err.Error(), 500) } } -// ReadNodeStatus populates the given NodeStatus struct with current -// values. -// -func ReadNodeStatus(st *NodeStatus) { +// populate the given NodeStatus struct with current values. +func readNodeStatus(st *NodeStatus) { vols := KeepVM.AllReadable() if cap(st.Volumes) < len(vols) { st.Volumes = make([]*VolumeStatus, len(vols)) @@ -231,9 +281,22 @@ func ReadNodeStatus(st *NodeStatus) { st.BufferPool.Alloc = bufs.Alloc() st.BufferPool.Cap = bufs.Cap() st.BufferPool.Len = bufs.Len() + st.PullQueue = getWorkQueueStatus(pullq) + st.TrashQueue = getWorkQueueStatus(trashq) runtime.ReadMemStats(&st.Memory) } +// return a WorkQueueStatus for the given queue. If q is nil (which +// should never happen except in test suites), return a zero status +// value instead of crashing. +func getWorkQueueStatus(q *WorkQueue) WorkQueueStatus { + if q == nil { + // This should only happen during tests. + return WorkQueueStatus{} + } + return q.Status() +} + // DeleteHandler processes DELETE requests. // // DELETE /{hash:[0-9a-f]{32} will delete the block with the specified hash @@ -271,7 +334,7 @@ func DeleteHandler(resp http.ResponseWriter, req *http.Request) { return } - if never_delete { + if neverDelete { http.Error(resp, MethodDisabledError.Error(), MethodDisabledError.HTTPCode) return } @@ -284,7 +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 @@ -308,7 +371,7 @@ func DeleteHandler(resp http.ResponseWriter, req *http.Request) { if body, err := json.Marshal(result); err == nil { resp.Write(body) } else { - log.Printf("json.Marshal: %s (result = %v)\n", err, result) + log.Printf("json.Marshal: %s (result = %v)", err, result) http.Error(resp, err.Error(), 500) } } @@ -347,11 +410,13 @@ func DeleteHandler(resp http.ResponseWriter, req *http.Request) { If the JSON unmarshalling fails, return 400 Bad Request. */ +// PullRequest consists of a block locator and an ordered list of servers type PullRequest struct { Locator string `json:"locator"` Servers []string `json:"servers"` } +// PullHandler processes "PUT /pull" requests for the data manager. func PullHandler(resp http.ResponseWriter, req *http.Request) { // Reject unauthorized requests. if !IsDataManagerToken(GetApiToken(req)) { @@ -381,11 +446,13 @@ func PullHandler(resp http.ResponseWriter, req *http.Request) { pullq.ReplaceQueue(plist) } +// TrashRequest consists of a block locator and it's Mtime type TrashRequest struct { Locator string `json:"locator"` BlockMtime int64 `json:"block_mtime"` } +// TrashHandler processes /trash requests. func TrashHandler(resp http.ResponseWriter, req *http.Request) { // Reject unauthorized requests. if !IsDataManagerToken(GetApiToken(req)) { @@ -415,7 +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 @@ -427,38 +540,20 @@ func TrashHandler(resp http.ResponseWriter, req *http.Request) { // 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". If -// the update_timestamp argument is true, GetBlock also updates the -// block's file modification time (for the sake of PutBlock, which -// must update the file's timestamp when the block already exists). -// -// 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, update_timestamp bool) ([]byte, error) { +func GetBlock(hash string, buf []byte, resp http.ResponseWriter) (int, error) { // Attempt to read the requested hash from a keep volume. - error_to_caller := NotFoundError - - var vols []Volume - if update_timestamp { - // Pointless to find the block on an unwritable volume - // because Touch() will fail -- this is as good as - // "not found" for purposes of callers who need to - // update_timestamp. - vols = KeepVM.AllWritable() - } else { - vols = KeepVM.AllReadable() - } + errorToCaller := NotFoundError - for _, vol := range vols { - buf, err := vol.Get(hash) + for _, vol := range KeepVM.AllReadable() { + size, err := vol.Get(hash, buf) if err != nil { // IsNotExist is an expected error and may be // ignored. All other errors are logged. In @@ -466,135 +561,152 @@ func GetBlock(hash string, update_timestamp bool) ([]byte, error) { // volumes. If all volumes report IsNotExist, // we return a NotFoundError. if !os.IsNotExist(err) { - log.Printf("GetBlock: reading %s: %s\n", hash, err) + log.Printf("%s: Get(%s): %s", vol, hash, err) } continue } // 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)\n", + log.Printf("%s: checksum mismatch for request %s (actual %s)", vol, hash, filehash) - error_to_caller = DiskHashError - bufs.Put(buf) + errorToCaller = DiskHashError continue } - if error_to_caller == DiskHashError { + if errorToCaller == DiskHashError { log.Printf("%s: checksum mismatch for request %s but a good copy was found on another volume and returned", vol, hash) } - if update_timestamp { - if err := vol.Touch(hash); err != nil { - error_to_caller = GenericError - log.Printf("%s: Touch %s failed: %s", - vol, hash, error_to_caller) - bufs.Put(buf) - continue - } - } - return buf, nil + return size, nil } - return nil, error_to_caller + return 0, errorToCaller } -/* PutBlock(block, hash) - Stores the BLOCK (identified by the content id HASH) in Keep. - - The MD5 checksum of the block must be identical to the content id HASH. - If not, an error is returned. - - PutBlock stores the BLOCK on the first Keep volume with free space. - A failure code is returned to the user only if all volumes fail. - - On success, PutBlock returns nil. - On failure, it returns a KeepError with one of the following codes: - - 500 Collision - A different block with the same hash already exists on this - Keep server. - 422 MD5Fail - The MD5 hash of the BLOCK does not match the argument HASH. - 503 Full - There was not enough space left in any Keep volume to store - the object. - 500 Fail - The object could not be stored for some other reason (e.g. - all writes failed). The text of the error message should - provide as much detail as possible. -*/ - -func PutBlock(block []byte, hash string) error { +// PutBlock Stores the BLOCK (identified by the content id HASH) in Keep. +// +// PutBlock(block, hash) +// Stores the BLOCK (identified by the content id HASH) in Keep. +// +// The MD5 checksum of the block must be identical to the content id HASH. +// If not, an error is returned. +// +// PutBlock stores the BLOCK on the first Keep volume with free space. +// A failure code is returned to the user only if all volumes fail. +// +// On success, PutBlock returns nil. +// On failure, it returns a KeepError with one of the following codes: +// +// 500 Collision +// A different block with the same hash already exists on this +// Keep server. +// 422 MD5Fail +// The MD5 hash of the BLOCK does not match the argument HASH. +// 503 Full +// There was not enough space left in any Keep volume to store +// the object. +// 500 Fail +// The object could not be stored for some other reason (e.g. +// all writes failed). The text of the error message should +// provide as much detail as possible. +// +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 - } - - // If we already have a block on disk under this identifier, return - // success (but check for MD5 collisions). While fetching the block, - // update its timestamp. - // The only errors that GetBlock can return are DiskHashError and NotFoundError. - // In either case, we want to write our new (good) block to disk, - // so there is nothing special to do if err != nil. - // - if oldblock, err := GetBlock(hash, true); err == nil { - defer bufs.Put(oldblock) - if bytes.Compare(block, oldblock) == 0 { - // The block already exists; return success. - return nil - } else { - return CollisionError - } + 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 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 // write did not succeed. Report the // error and continue trying. allFull = false - log.Printf("%s: Write(%s): %s\n", vol, hash, err) + log.Printf("%s: Write(%s): %s", vol, hash, err) } } if allFull { log.Print("All volumes are full.") - return FullError - } else { - // Already logged the non-full errors. - return GenericError + return 0, FullError + } + // Already logged the non-full errors. + return 0, GenericError +} + +// 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 { + // Stop if we have a block with same hash but + // different content. (It will be impossible + // to tell which one is wanted if we have + // both, so there's no point writing it even + // on a different volume.) + log.Printf("%s: Compare(%s): %s", vol, hash, err) + return 0, err + } else if os.IsNotExist(err) { + // Block does not exist. This is the only + // "normal" error: we don't log anything. + continue + } else if err != nil { + // Couldn't open file, data is corrupt on + // disk, etc.: log this abnormal condition, + // and try the next volume. + log.Printf("%s: Compare(%s): %s", vol, hash, err) + continue + } + if err := vol.Touch(hash); err != nil { + log.Printf("%s: Touch %s failed: %s", vol, hash, err) + bestErr = err + continue + } + // Compare and Touch both worked --> done. + return vol.Replication(), nil } + return 0, bestErr } var validLocatorRe = regexp.MustCompile(`^[0-9a-f]{32}$`) -// IsValidLocator -// Return true if the specified string is a valid Keep locator. -// When Keep is extended to support hash types other than MD5, -// this should be updated to cover those as well. +// IsValidLocator returns true if the specified string is a valid Keep locator. +// When Keep is extended to support hash types other than MD5, +// this should be updated to cover those as well. // func IsValidLocator(loc string) bool { return validLocatorRe.MatchString(loc) @@ -615,36 +727,36 @@ func GetApiToken(req *http.Request) string { } // IsExpired returns true if the given Unix timestamp (expressed as a -// hexadecimal string) is in the past, or if timestamp_hex cannot be +// hexadecimal string) is in the past, or if timestampHex cannot be // parsed as a hexadecimal string. -func IsExpired(timestamp_hex string) bool { - ts, err := strconv.ParseInt(timestamp_hex, 16, 0) +func IsExpired(timestampHex string) bool { + ts, err := strconv.ParseInt(timestampHex, 16, 0) if err != nil { - log.Printf("IsExpired: %s\n", err) + log.Printf("IsExpired: %s", err) return true } return time.Unix(ts, 0).Before(time.Now()) } -// CanDelete returns true if the user identified by api_token is +// CanDelete returns true if the user identified by apiToken is // allowed to delete blocks. -func CanDelete(api_token string) bool { - if api_token == "" { +func CanDelete(apiToken string) bool { + if apiToken == "" { return false } // Blocks may be deleted only when Keep has been configured with a // data manager. - if IsDataManagerToken(api_token) { + if IsDataManagerToken(apiToken) { return true } - // TODO(twp): look up api_token with the API server + // TODO(twp): look up apiToken with the API server // return true if is_admin is true and if the token // has unlimited scope return false } -// IsDataManagerToken returns true if api_token represents the data +// IsDataManagerToken returns true if apiToken represents the data // manager's token. -func IsDataManagerToken(api_token string) bool { - return data_manager_token != "" && api_token == data_manager_token +func IsDataManagerToken(apiToken string) bool { + return dataManagerToken != "" && apiToken == dataManagerToken }