16306: Merge branch 'master'
[arvados.git] / sdk / go / keepclient / keepclient.go
index b56cc7f724b3ba64ee26033f5ddd4b6f888f2422..b18d7e046404c34b0f1486aadfb667a2fe1e01d7 100644 (file)
@@ -1,3 +1,7 @@
+// Copyright (C) The Arvados Authors. All rights reserved.
+//
+// SPDX-License-Identifier: Apache-2.0
+
 /* Provides low-level Get/Put primitives for accessing Arvados Keep blocks. */
 package keepclient
 
@@ -8,19 +12,34 @@ import (
        "fmt"
        "io"
        "io/ioutil"
+       "net"
        "net/http"
        "regexp"
        "strconv"
        "strings"
        "sync"
+       "time"
 
-       "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
-       "git.curoverse.com/arvados.git/sdk/go/streamer"
+       "git.arvados.org/arvados.git/sdk/go/arvadosclient"
+       "git.arvados.org/arvados.git/sdk/go/asyncbuf"
+       "git.arvados.org/arvados.git/sdk/go/httpserver"
 )
 
 // A Keep "block" is 64MB.
 const BLOCKSIZE = 64 * 1024 * 1024
 
+var (
+       DefaultRequestTimeout      = 20 * time.Second
+       DefaultConnectTimeout      = 2 * time.Second
+       DefaultTLSHandshakeTimeout = 4 * time.Second
+       DefaultKeepAlive           = 180 * time.Second
+
+       DefaultProxyRequestTimeout      = 300 * time.Second
+       DefaultProxyConnectTimeout      = 30 * time.Second
+       DefaultProxyTLSHandshakeTimeout = 10 * time.Second
+       DefaultProxyKeepAlive           = 120 * time.Second
+)
+
 // Error interface with an error and boolean indicating whether the error is temporary
 type Error interface {
        error
@@ -74,29 +93,36 @@ type HTTPClient interface {
 type KeepClient struct {
        Arvados            *arvadosclient.ArvadosClient
        Want_replicas      int
-       localRoots         *map[string]string
-       writableLocalRoots *map[string]string
-       gatewayRoots       *map[string]string
+       localRoots         map[string]string
+       writableLocalRoots map[string]string
+       gatewayRoots       map[string]string
        lock               sync.RWMutex
-       Client             HTTPClient
+       HTTPClient         HTTPClient
        Retries            int
        BlockCache         *BlockCache
+       RequestID          string
+       StorageClasses     []string
 
        // set to 1 if all writable services are of disk type, otherwise 0
        replicasPerService int
 
        // Any non-disk typed services found in the list of keepservers?
        foundNonDiskSvc bool
+
+       // Disable automatic discovery of keep services
+       disableDiscovery bool
 }
 
-// MakeKeepClient creates a new KeepClient by contacting the API server to discover Keep servers.
+// MakeKeepClient creates a new KeepClient, calls
+// DiscoverKeepServices(), and returns when the client is ready to
+// use.
 func MakeKeepClient(arv *arvadosclient.ArvadosClient) (*KeepClient, error) {
        kc := New(arv)
-       return kc, kc.DiscoverKeepServers()
+       return kc, kc.discoverServices()
 }
 
-// New func creates a new KeepClient struct.
-// This func does not discover keep servers. It is the caller's responsibility.
+// New creates a new KeepClient. Service discovery will occur on the
+// next read/write operation.
 func New(arv *arvadosclient.ArvadosClient) *KeepClient {
        defaultReplicationLevel := 2
        value, err := arv.Discovery("defaultCollectionReplication")
@@ -106,15 +132,11 @@ func New(arv *arvadosclient.ArvadosClient) *KeepClient {
                        defaultReplicationLevel = int(v)
                }
        }
-
-       kc := &KeepClient{
+       return &KeepClient{
                Arvados:       arv,
                Want_replicas: defaultReplicationLevel,
-               Client: &http.Client{Transport: &http.Transport{
-                       TLSClientConfig: arvadosclient.MakeTLSConfig(arv.ApiInsecure)}},
-               Retries: 2,
+               Retries:       2,
        }
-       return kc
 }
 
 // Put a block given the block hash, a reader, and the number of bytes
@@ -137,10 +159,12 @@ func (kc *KeepClient) PutHR(hash string, r io.Reader, dataBytes int64) (string,
                bufsize = BLOCKSIZE
        }
 
-       t := streamer.AsyncStreamFromReader(bufsize, HashCheckingReader{r, md5.New(), hash})
-       defer t.Close()
-
-       return kc.putReplicas(hash, t, dataBytes)
+       buf := asyncbuf.NewBuffer(make([]byte, 0, bufsize))
+       go func() {
+               _, err := io.Copy(buf, HashCheckingReader{r, md5.New(), hash})
+               buf.CloseWithError(err)
+       }()
+       return kc.putReplicas(hash, buf.NewReader, dataBytes)
 }
 
 // PutHB writes a block to Keep. The hash of the bytes is given in
@@ -148,9 +172,8 @@ func (kc *KeepClient) PutHR(hash string, r io.Reader, dataBytes int64) (string,
 //
 // Return values are the same as for PutHR.
 func (kc *KeepClient) PutHB(hash string, buf []byte) (string, int, error) {
-       t := streamer.AsyncStreamFromSlice(buf)
-       defer t.Close()
-       return kc.putReplicas(hash, t, int64(len(buf)))
+       newReader := func() io.Reader { return bytes.NewBuffer(buf) }
+       return kc.putReplicas(hash, newReader, int64(len(buf)))
 }
 
 // PutB writes a block to Keep. It computes the hash itself.
@@ -175,9 +198,20 @@ func (kc *KeepClient) PutR(r io.Reader) (locator string, replicas int, err error
        }
 }
 
-func (kc *KeepClient) getOrHead(method string, locator string) (io.ReadCloser, int64, string, error) {
+func (kc *KeepClient) getOrHead(method string, locator string, header http.Header) (io.ReadCloser, int64, string, http.Header, error) {
        if strings.HasPrefix(locator, "d41d8cd98f00b204e9800998ecf8427e+0") {
-               return ioutil.NopCloser(bytes.NewReader(nil)), 0, "", nil
+               return ioutil.NopCloser(bytes.NewReader(nil)), 0, "", nil, nil
+       }
+
+       reqid := kc.getRequestID()
+
+       var expectLength int64
+       if parts := strings.SplitN(locator, "+", 3); len(parts) < 2 {
+               expectLength = -1
+       } else if n, err := strconv.ParseInt(parts[1], 10, 64); err != nil {
+               expectLength = -1
+       } else {
+               expectLength = n
        }
 
        var errs []string
@@ -203,14 +237,24 @@ func (kc *KeepClient) getOrHead(method string, locator string) (io.ReadCloser, i
                                errs = append(errs, fmt.Sprintf("%s: %v", url, err))
                                continue
                        }
-                       req.Header.Add("Authorization", fmt.Sprintf("OAuth2 %s", kc.Arvados.ApiToken))
-                       resp, err := kc.Client.Do(req)
+                       for k, v := range header {
+                               req.Header[k] = append([]string(nil), v...)
+                       }
+                       if req.Header.Get("Authorization") == "" {
+                               req.Header.Set("Authorization", "OAuth2 "+kc.Arvados.ApiToken)
+                       }
+                       if req.Header.Get("X-Request-Id") == "" {
+                               req.Header.Set("X-Request-Id", reqid)
+                       }
+                       resp, err := kc.httpClient().Do(req)
                        if err != nil {
                                // Probably a network error, may be transient,
                                // can try again.
                                errs = append(errs, fmt.Sprintf("%s: %v", url, err))
                                retryList = append(retryList, host)
-                       } else if resp.StatusCode != http.StatusOK {
+                               continue
+                       }
+                       if resp.StatusCode != http.StatusOK {
                                var respbody []byte
                                respbody, _ = ioutil.ReadAll(&io.LimitedReader{R: resp.Body, N: 4096})
                                resp.Body.Close()
@@ -227,20 +271,29 @@ func (kc *KeepClient) getOrHead(method string, locator string) (io.ReadCloser, i
                                } else if resp.StatusCode == 404 {
                                        count404++
                                }
-                       } else {
-                               // Success.
-                               if method == "GET" {
-                                       return HashCheckingReader{
-                                               Reader: resp.Body,
-                                               Hash:   md5.New(),
-                                               Check:  locator[0:32],
-                                       }, resp.ContentLength, url, nil
-                               } else {
+                               continue
+                       }
+                       if expectLength < 0 {
+                               if resp.ContentLength < 0 {
                                        resp.Body.Close()
-                                       return nil, resp.ContentLength, url, nil
+                                       return nil, 0, "", nil, fmt.Errorf("error reading %q: no size hint, no Content-Length header in response", locator)
                                }
+                               expectLength = resp.ContentLength
+                       } else if resp.ContentLength >= 0 && expectLength != resp.ContentLength {
+                               resp.Body.Close()
+                               return nil, 0, "", nil, fmt.Errorf("error reading %q: size hint %d != Content-Length %d", locator, expectLength, resp.ContentLength)
+                       }
+                       // Success
+                       if method == "GET" {
+                               return HashCheckingReader{
+                                       Reader: resp.Body,
+                                       Hash:   md5.New(),
+                                       Check:  locator[0:32],
+                               }, expectLength, url, resp.Header, nil
+                       } else {
+                               resp.Body.Close()
+                               return nil, expectLength, url, resp.Header, nil
                        }
-
                }
                serversToTry = retryList
        }
@@ -255,7 +308,29 @@ func (kc *KeepClient) getOrHead(method string, locator string) (io.ReadCloser, i
                        isTemp: len(serversToTry) > 0,
                }}
        }
-       return nil, 0, "", err
+       return nil, 0, "", nil, err
+}
+
+// LocalLocator returns a locator equivalent to the one supplied, but
+// with a valid signature from the local cluster. If the given locator
+// already has a local signature, it is returned unchanged.
+func (kc *KeepClient) LocalLocator(locator string) (string, error) {
+       if !strings.Contains(locator, "+R") {
+               // Either it has +A, or it's unsigned and we assume
+               // it's a local locator on a site with signatures
+               // disabled.
+               return locator, nil
+       }
+       sighdr := fmt.Sprintf("local, time=%s", time.Now().UTC().Format(time.RFC3339))
+       _, _, url, hdr, err := kc.getOrHead("HEAD", locator, http.Header{"X-Keep-Signature": []string{sighdr}})
+       if err != nil {
+               return "", err
+       }
+       loc := hdr.Get("X-Keep-Locator")
+       if loc == "" {
+               return "", fmt.Errorf("missing X-Keep-Locator header in HEAD response from %s", url)
+       }
+       return loc, nil
 }
 
 // Get() retrieves a block, given a locator. Returns a reader, the
@@ -266,7 +341,14 @@ func (kc *KeepClient) getOrHead(method string, locator string) (io.ReadCloser, i
 // reader returned by this method will return a BadChecksum error
 // instead of EOF.
 func (kc *KeepClient) Get(locator string) (io.ReadCloser, int64, string, error) {
-       return kc.getOrHead("GET", locator)
+       rdr, size, url, _, err := kc.getOrHead("GET", locator, nil)
+       return rdr, size, url, err
+}
+
+// ReadAt() retrieves a portion of block from the cache if it's
+// present, otherwise from the network.
+func (kc *KeepClient) ReadAt(locator string, p []byte, off int) (int, error) {
+       return kc.cache().ReadAt(kc, locator, p, off)
 }
 
 // Ask() verifies that a block with the given hash is available and
@@ -277,7 +359,7 @@ func (kc *KeepClient) Get(locator string) (io.ReadCloser, int64, string, error)
 // Returns the data size (content length) reported by the Keep service
 // and the URI reporting the data size.
 func (kc *KeepClient) Ask(locator string) (int64, string, error) {
-       _, size, url, err := kc.getOrHead("HEAD", locator)
+       _, size, url, _, err := kc.getOrHead("HEAD", locator, nil)
        return size, url, err
 }
 
@@ -304,8 +386,9 @@ func (kc *KeepClient) GetIndex(keepServiceUUID, prefix string) (io.Reader, error
                return nil, err
        }
 
-       req.Header.Add("Authorization", fmt.Sprintf("OAuth2 %s", kc.Arvados.ApiToken))
-       resp, err := kc.Client.Do(req)
+       req.Header.Add("Authorization", "OAuth2 "+kc.Arvados.ApiToken)
+       req.Header.Set("X-Request-Id", kc.getRequestID())
+       resp, err := kc.httpClient().Do(req)
        if err != nil {
                return nil, err
        }
@@ -336,55 +419,47 @@ func (kc *KeepClient) GetIndex(keepServiceUUID, prefix string) (io.Reader, error
 // LocalRoots() returns the map of local (i.e., disk and proxy) Keep
 // services: uuid -> baseURI.
 func (kc *KeepClient) LocalRoots() map[string]string {
+       kc.discoverServices()
        kc.lock.RLock()
        defer kc.lock.RUnlock()
-       return *kc.localRoots
+       return kc.localRoots
 }
 
 // GatewayRoots() returns the map of Keep remote gateway services:
 // uuid -> baseURI.
 func (kc *KeepClient) GatewayRoots() map[string]string {
+       kc.discoverServices()
        kc.lock.RLock()
        defer kc.lock.RUnlock()
-       return *kc.gatewayRoots
+       return kc.gatewayRoots
 }
 
 // WritableLocalRoots() returns the map of writable local Keep services:
 // uuid -> baseURI.
 func (kc *KeepClient) WritableLocalRoots() map[string]string {
+       kc.discoverServices()
        kc.lock.RLock()
        defer kc.lock.RUnlock()
-       return *kc.writableLocalRoots
+       return kc.writableLocalRoots
 }
 
-// SetServiceRoots updates the localRoots and gatewayRoots maps,
-// without risk of disrupting operations that are already in progress.
+// SetServiceRoots disables service discovery and updates the
+// localRoots and gatewayRoots maps, without disrupting operations
+// that are already in progress.
 //
-// The KeepClient makes its own copy of the supplied maps, so the
-// caller can reuse/modify them after SetServiceRoots returns, but
-// they should not be modified by any other goroutine while
-// SetServiceRoots is running.
-func (kc *KeepClient) SetServiceRoots(newLocals, newWritableLocals, newGateways map[string]string) {
-       locals := make(map[string]string)
-       for uuid, root := range newLocals {
-               locals[uuid] = root
-       }
-
-       writables := make(map[string]string)
-       for uuid, root := range newWritableLocals {
-               writables[uuid] = root
-       }
-
-       gateways := make(map[string]string)
-       for uuid, root := range newGateways {
-               gateways[uuid] = root
-       }
+// The supplied maps must not be modified after calling
+// SetServiceRoots.
+func (kc *KeepClient) SetServiceRoots(locals, writables, gateways map[string]string) {
+       kc.disableDiscovery = true
+       kc.setServiceRoots(locals, writables, gateways)
+}
 
+func (kc *KeepClient) setServiceRoots(locals, writables, gateways map[string]string) {
        kc.lock.Lock()
        defer kc.lock.Unlock()
-       kc.localRoots = &locals
-       kc.writableLocalRoots = &writables
-       kc.gatewayRoots = &gateways
+       kc.localRoots = locals
+       kc.writableLocalRoots = writables
+       kc.gatewayRoots = gateways
 }
 
 // getSortedRoots returns a list of base URIs of Keep services, in the
@@ -423,6 +498,89 @@ func (kc *KeepClient) cache() *BlockCache {
        }
 }
 
+func (kc *KeepClient) ClearBlockCache() {
+       kc.cache().Clear()
+}
+
+var (
+       // There are four global http.Client objects for the four
+       // possible permutations of TLS behavior (verify/skip-verify)
+       // and timeout settings (proxy/non-proxy).
+       defaultClient = map[bool]map[bool]HTTPClient{
+               // defaultClient[false] is used for verified TLS reqs
+               false: {},
+               // defaultClient[true] is used for unverified
+               // (insecure) TLS reqs
+               true: {},
+       }
+       defaultClientMtx sync.Mutex
+)
+
+// httpClient returns the HTTPClient field if it's not nil, otherwise
+// whichever of the four global http.Client objects is suitable for
+// the current environment (i.e., TLS verification on/off, keep
+// services are/aren't proxies).
+func (kc *KeepClient) httpClient() HTTPClient {
+       if kc.HTTPClient != nil {
+               return kc.HTTPClient
+       }
+       defaultClientMtx.Lock()
+       defer defaultClientMtx.Unlock()
+       if c, ok := defaultClient[kc.Arvados.ApiInsecure][kc.foundNonDiskSvc]; ok {
+               return c
+       }
+
+       var requestTimeout, connectTimeout, keepAlive, tlsTimeout time.Duration
+       if kc.foundNonDiskSvc {
+               // Use longer timeouts when connecting to a proxy,
+               // because this usually means the intervening network
+               // is slower.
+               requestTimeout = DefaultProxyRequestTimeout
+               connectTimeout = DefaultProxyConnectTimeout
+               tlsTimeout = DefaultProxyTLSHandshakeTimeout
+               keepAlive = DefaultProxyKeepAlive
+       } else {
+               requestTimeout = DefaultRequestTimeout
+               connectTimeout = DefaultConnectTimeout
+               tlsTimeout = DefaultTLSHandshakeTimeout
+               keepAlive = DefaultKeepAlive
+       }
+
+       c := &http.Client{
+               Timeout: requestTimeout,
+               // It's not safe to copy *http.DefaultTransport
+               // because it has a mutex (which might be locked)
+               // protecting a private map (which might not be nil).
+               // So we build our own, using the Go 1.12 default
+               // values, ignoring any changes the application has
+               // made to http.DefaultTransport.
+               Transport: &http.Transport{
+                       DialContext: (&net.Dialer{
+                               Timeout:   connectTimeout,
+                               KeepAlive: keepAlive,
+                               DualStack: true,
+                       }).DialContext,
+                       MaxIdleConns:          100,
+                       IdleConnTimeout:       90 * time.Second,
+                       TLSHandshakeTimeout:   tlsTimeout,
+                       ExpectContinueTimeout: 1 * time.Second,
+                       TLSClientConfig:       arvadosclient.MakeTLSConfig(kc.Arvados.ApiInsecure),
+               },
+       }
+       defaultClient[kc.Arvados.ApiInsecure][kc.foundNonDiskSvc] = c
+       return c
+}
+
+var reqIDGen = httpserver.IDGenerator{Prefix: "req-"}
+
+func (kc *KeepClient) getRequestID() string {
+       if kc.RequestID != "" {
+               return kc.RequestID
+       } else {
+               return reqIDGen.Next()
+       }
+}
+
 type Locator struct {
        Hash  string
        Size  int      // -1 if data size is not known