services/keep-web/s3.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package keepweb
   6
   7 import (
   8         "crypto/hmac"
   9         "crypto/sha256"
  10         "encoding/base64"
  11         "encoding/json"
  12         "encoding/xml"
  13         "errors"
  14         "fmt"
  15         "hash"
  16         "io"
  17         "net/http"
  18         "net/textproto"
  19         "net/url"
  20         "os"
  21         "path/filepath"
  22         "regexp"
  23         "sort"
  24         "strconv"
  25         "strings"
  26         "time"
  27
  28         "git.arvados.org/arvados.git/sdk/go/arvados"
  29         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
  30         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  31         "git.arvados.org/arvados.git/sdk/go/keepclient"
  32         "github.com/AdRoll/goamz/s3"
  33 )
  34
  35 const (
  36         s3MaxKeys       = 1000
  37         s3SignAlgorithm = "AWS4-HMAC-SHA256"
  38         s3MaxClockSkew  = 5 * time.Minute
  39 )
  40
  41 type commonPrefix struct {
  42         Prefix string
  43 }
  44
  45 type listV1Resp struct {
  46         XMLName string `xml:"http://s3.amazonaws.com/doc/2006-03-01/ ListBucketResult"`
  47         s3.ListResp
  48         // s3.ListResp marshals an empty tag when
  49         // CommonPrefixes is nil, which confuses some clients.
  50         // Fix by using this nested struct instead.
  51         CommonPrefixes []commonPrefix
  52         // Similarly, we need omitempty here, because an empty
  53         // tag confuses some clients (e.g.,
  54         // github.com/aws/aws-sdk-net never terminates its
  55         // paging loop).
  56         NextMarker string `xml:"NextMarker,omitempty"`
  57         // ListObjectsV2 has a KeyCount response field.
  58         KeyCount int
  59 }
  60
  61 type listV2Resp struct {
  62         XMLName               string `xml:"http://s3.amazonaws.com/doc/2006-03-01/ ListBucketResult"`
  63         IsTruncated           bool
  64         Contents              []s3.Key
  65         Name                  string
  66         Prefix                string
  67         Delimiter             string
  68         MaxKeys               int
  69         CommonPrefixes        []commonPrefix
  70         EncodingType          string `xml:",omitempty"`
  71         KeyCount              int
  72         ContinuationToken     string `xml:",omitempty"`
  73         NextContinuationToken string `xml:",omitempty"`
  74         StartAfter            string `xml:",omitempty"`
  75 }
  76
  77 func hmacstring(msg string, key []byte) []byte {
  78         h := hmac.New(sha256.New, key)
  79         io.WriteString(h, msg)
  80         return h.Sum(nil)
  81 }
  82
  83 func hashdigest(h hash.Hash, payload string) string {
  84         io.WriteString(h, payload)
  85         return fmt.Sprintf("%x", h.Sum(nil))
  86 }
  87
  88 // Signing key for given secret key and request attrs.
  89 func s3signatureKey(key, datestamp, regionName, serviceName string) []byte {
  90         return hmacstring("aws4_request",
  91                 hmacstring(serviceName,
  92                         hmacstring(regionName,
  93                                 hmacstring(datestamp, []byte("AWS4"+key)))))
  94 }
  95
  96 // Canonical query string for S3 V4 signature: sorted keys, spaces
  97 // escaped as %20 instead of +, keyvalues joined with &.
  98 func s3querystring(u *url.URL) string {
  99         keys := make([]string, 0, len(u.Query()))
 100         values := make(map[string]string, len(u.Query()))
 101         for k, vs := range u.Query() {
 102                 k = strings.Replace(url.QueryEscape(k), "+", "%20", -1)
 103                 keys = append(keys, k)
 104                 for _, v := range vs {
 105                         v = strings.Replace(url.QueryEscape(v), "+", "%20", -1)
 106                         if values[k] != "" {
 107                                 values[k] += "&"
 108                         }
 109                         values[k] += k + "=" + v
 110                 }
 111         }
 112         sort.Strings(keys)
 113         for i, k := range keys {
 114                 keys[i] = values[k]
 115         }
 116         return strings.Join(keys, "&")
 117 }
 118
 119 var reMultipleSlashChars = regexp.MustCompile(`//+`)
 120
 121 func s3stringToSign(alg, scope, signedHeaders string, r *http.Request) (string, error) {
 122         timefmt, timestr := "20060102T150405Z", r.Header.Get("X-Amz-Date")
 123         if timestr == "" {
 124                 timefmt, timestr = time.RFC1123, r.Header.Get("Date")
 125         }
 126         t, err := time.Parse(timefmt, timestr)
 127         if err != nil {
 128                 return "", fmt.Errorf("invalid timestamp %q: %s", timestr, err)
 129         }
 130         if skew := time.Now().Sub(t); skew < -s3MaxClockSkew || skew > s3MaxClockSkew {
 131                 return "", errors.New("exceeded max clock skew")
 132         }
 133
 134         var canonicalHeaders string
 135         for _, h := range strings.Split(signedHeaders, ";") {
 136                 if h == "host" {
 137                         canonicalHeaders += h + ":" + r.Host + "\n"
 138                 } else {
 139                         canonicalHeaders += h + ":" + r.Header.Get(h) + "\n"
 140                 }
 141         }
 142
 143         normalizedPath := normalizePath(r.URL.Path)
 144         ctxlog.FromContext(r.Context()).Debugf("normalizedPath %q", normalizedPath)
 145         canonicalRequest := fmt.Sprintf("%s\n%s\n%s\n%s\n%s\n%s", r.Method, normalizedPath, s3querystring(r.URL), canonicalHeaders, signedHeaders, r.Header.Get("X-Amz-Content-Sha256"))
 146         ctxlog.FromContext(r.Context()).Debugf("s3stringToSign: canonicalRequest %s", canonicalRequest)
 147         return fmt.Sprintf("%s\n%s\n%s\n%s", alg, r.Header.Get("X-Amz-Date"), scope, hashdigest(sha256.New(), canonicalRequest)), nil
 148 }
 149
 150 func normalizePath(s string) string {
 151         // (url.URL).EscapedPath() would be incorrect here. AWS
 152         // documentation specifies the URL path should be normalized
 153         // according to RFC 3986, i.e., unescaping ALPHA / DIGIT / "-"
 154         // / "." / "_" / "~". The implication is that everything other
 155         // than those chars (and "/") _must_ be percent-encoded --
 156         // even chars like ";" and "," that are not normally
 157         // percent-encoded in paths.
 158         out := ""
 159         for _, c := range []byte(reMultipleSlashChars.ReplaceAllString(s, "/")) {
 160                 if (c >= 'a' && c <= 'z') ||
 161                         (c >= 'A' && c <= 'Z') ||
 162                         (c >= '0' && c <= '9') ||
 163                         c == '-' ||
 164                         c == '.' ||
 165                         c == '_' ||
 166                         c == '~' ||
 167                         c == '/' {
 168                         out += string(c)
 169                 } else {
 170                         out += fmt.Sprintf("%%%02X", c)
 171                 }
 172         }
 173         return out
 174 }
 175
 176 func s3signature(secretKey, scope, signedHeaders, stringToSign string) (string, error) {
 177         // scope is {datestamp}/{region}/{service}/aws4_request
 178         drs := strings.Split(scope, "/")
 179         if len(drs) != 4 {
 180                 return "", fmt.Errorf("invalid scope %q", scope)
 181         }
 182         key := s3signatureKey(secretKey, drs[0], drs[1], drs[2])
 183         return hashdigest(hmac.New(sha256.New, key), stringToSign), nil
 184 }
 185
 186 var v2tokenUnderscore = regexp.MustCompile(`^v2_[a-z0-9]{5}-gj3su-[a-z0-9]{15}_`)
 187
 188 func unescapeKey(key string) string {
 189         if v2tokenUnderscore.MatchString(key) {
 190                 // Entire Arvados token, with "/" replaced by "_" to
 191                 // avoid colliding with the Authorization header
 192                 // format.
 193                 return strings.Replace(key, "_", "/", -1)
 194         } else if s, err := url.PathUnescape(key); err == nil {
 195                 return s
 196         } else {
 197                 return key
 198         }
 199 }
 200
 201 // checks3signature verifies the given S3 V4 signature and returns the
 202 // Arvados token that corresponds to the given accessKey. An error is
 203 // returned if accessKey is not a valid token UUID or the signature
 204 // does not match.
 205 func (h *handler) checks3signature(r *http.Request) (string, error) {
 206         var key, scope, signedHeaders, signature string
 207         authstring := strings.TrimPrefix(r.Header.Get("Authorization"), s3SignAlgorithm+" ")
 208         for _, cmpt := range strings.Split(authstring, ",") {
 209                 cmpt = strings.TrimSpace(cmpt)
 210                 split := strings.SplitN(cmpt, "=", 2)
 211                 switch {
 212                 case len(split) != 2:
 213                         // (?) ignore
 214                 case split[0] == "Credential":
 215                         keyandscope := strings.SplitN(split[1], "/", 2)
 216                         if len(keyandscope) == 2 {
 217                                 key, scope = keyandscope[0], keyandscope[1]
 218                         }
 219                 case split[0] == "SignedHeaders":
 220                         signedHeaders = split[1]
 221                 case split[0] == "Signature":
 222                         signature = split[1]
 223                 }
 224         }
 225
 226         client := (&arvados.Client{
 227                 APIHost:  h.Cluster.Services.Controller.ExternalURL.Host,
 228                 Insecure: h.Cluster.TLS.Insecure,
 229         }).WithRequestID(r.Header.Get("X-Request-Id"))
 230         var aca arvados.APIClientAuthorization
 231         var secret string
 232         var err error
 233         if len(key) == 27 && key[5:12] == "-gj3su-" {
 234                 // Access key is the UUID of an Arvados token, secret
 235                 // key is the secret part.
 236                 ctx := arvados.ContextWithAuthorization(r.Context(), "Bearer "+h.Cluster.SystemRootToken)
 237                 err = client.RequestAndDecodeContext(ctx, &aca, "GET", "arvados/v1/api_client_authorizations/"+key, nil, nil)
 238                 secret = aca.APIToken
 239         } else {
 240                 // Access key and secret key are both an entire
 241                 // Arvados token or OIDC access token.
 242                 ctx := arvados.ContextWithAuthorization(r.Context(), "Bearer "+unescapeKey(key))
 243                 err = client.RequestAndDecodeContext(ctx, &aca, "GET", "arvados/v1/api_client_authorizations/current", nil, nil)
 244                 secret = key
 245         }
 246         if err != nil {
 247                 ctxlog.FromContext(r.Context()).WithError(err).WithField("UUID", key).Info("token lookup failed")
 248                 return "", errors.New("invalid access key")
 249         }
 250         stringToSign, err := s3stringToSign(s3SignAlgorithm, scope, signedHeaders, r)
 251         if err != nil {
 252                 return "", err
 253         }
 254         expect, err := s3signature(secret, scope, signedHeaders, stringToSign)
 255         if err != nil {
 256                 return "", err
 257         } else if expect != signature {
 258                 return "", fmt.Errorf("signature does not match (scope %q signedHeaders %q stringToSign %q)", scope, signedHeaders, stringToSign)
 259         }
 260         return aca.TokenV2(), nil
 261 }
 262
 263 func s3ErrorResponse(w http.ResponseWriter, s3code string, message string, resource string, code int) {
 264         w.Header().Set("Content-Type", "application/xml")
 265         w.Header().Set("X-Content-Type-Options", "nosniff")
 266         w.WriteHeader(code)
 267         var errstruct struct {
 268                 Code      string
 269                 Message   string
 270                 Resource  string
 271                 RequestId string
 272         }
 273         errstruct.Code = s3code
 274         errstruct.Message = message
 275         errstruct.Resource = resource
 276         errstruct.RequestId = ""
 277         enc := xml.NewEncoder(w)
 278         fmt.Fprint(w, xml.Header)
 279         enc.EncodeElement(errstruct, xml.StartElement{Name: xml.Name{Local: "Error"}})
 280 }
 281
 282 var NoSuchKey = "NoSuchKey"
 283 var NoSuchBucket = "NoSuchBucket"
 284 var InvalidArgument = "InvalidArgument"
 285 var InternalError = "InternalError"
 286 var UnauthorizedAccess = "UnauthorizedAccess"
 287 var InvalidRequest = "InvalidRequest"
 288 var SignatureDoesNotMatch = "SignatureDoesNotMatch"
 289
 290 var reRawQueryIndicatesAPI = regexp.MustCompile(`^[a-z]+(&|$)`)
 291
 292 // serveS3 handles r and returns true if r is a request from an S3
 293 // client, otherwise it returns false.
 294 func (h *handler) serveS3(w http.ResponseWriter, r *http.Request) bool {
 295         var token string
 296         if auth := r.Header.Get("Authorization"); strings.HasPrefix(auth, "AWS ") {
 297                 split := strings.SplitN(auth[4:], ":", 2)
 298                 if len(split) < 2 {
 299                         s3ErrorResponse(w, InvalidRequest, "malformed Authorization header", r.URL.Path, http.StatusUnauthorized)
 300                         return true
 301                 }
 302                 token = unescapeKey(split[0])
 303         } else if strings.HasPrefix(auth, s3SignAlgorithm+" ") {
 304                 t, err := h.checks3signature(r)
 305                 if err != nil {
 306                         s3ErrorResponse(w, SignatureDoesNotMatch, "signature verification failed: "+err.Error(), r.URL.Path, http.StatusForbidden)
 307                         return true
 308                 }
 309                 token = t
 310         } else {
 311                 return false
 312         }
 313
 314         var err error
 315         var fs arvados.CustomFileSystem
 316         var arvclient *arvadosclient.ArvadosClient
 317         if r.Method == http.MethodGet || r.Method == http.MethodHead {
 318                 // Use a single session (cached FileSystem) across
 319                 // multiple read requests.
 320                 var sess *cachedSession
 321                 fs, sess, err = h.Cache.GetSession(token)
 322                 if err != nil {
 323                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 324                         return true
 325                 }
 326                 arvclient = sess.arvadosclient
 327         } else {
 328                 // Create a FileSystem for this request, to avoid
 329                 // exposing incomplete write operations to concurrent
 330                 // requests.
 331                 var kc *keepclient.KeepClient
 332                 var release func()
 333                 var client *arvados.Client
 334                 arvclient, kc, client, release, err = h.getClients(r.Header.Get("X-Request-Id"), token)
 335                 if err != nil {
 336                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 337                         return true
 338                 }
 339                 defer release()
 340                 fs = client.SiteFileSystem(kc)
 341                 fs.ForwardSlashNameSubstitution(h.Cluster.Collections.ForwardSlashNameSubstitution)
 342         }
 343
 344         var objectNameGiven bool
 345         var bucketName string
 346         fspath := "/by_id"
 347         if id := arvados.CollectionIDFromDNSName(r.Host); id != "" {
 348                 fspath += "/" + id
 349                 bucketName = id
 350                 objectNameGiven = strings.Count(strings.TrimSuffix(r.URL.Path, "/"), "/") > 0
 351         } else {
 352                 bucketName = strings.SplitN(strings.TrimPrefix(r.URL.Path, "/"), "/", 2)[0]
 353                 objectNameGiven = strings.Count(strings.TrimSuffix(r.URL.Path, "/"), "/") > 1
 354         }
 355         fspath += reMultipleSlashChars.ReplaceAllString(r.URL.Path, "/")
 356
 357         switch {
 358         case r.Method == http.MethodGet && !objectNameGiven:
 359                 // Path is "/{uuid}" or "/{uuid}/", has no object name
 360                 if _, ok := r.URL.Query()["versioning"]; ok {
 361                         // GetBucketVersioning
 362                         w.Header().Set("Content-Type", "application/xml")
 363                         io.WriteString(w, xml.Header)
 364                         fmt.Fprintln(w, `<VersioningConfiguration xmlns="http://s3.amazonaws.com/doc/2006-03-01/"/>`)
 365                 } else if _, ok = r.URL.Query()["location"]; ok {
 366                         // GetBucketLocation
 367                         w.Header().Set("Content-Type", "application/xml")
 368                         io.WriteString(w, xml.Header)
 369                         fmt.Fprintln(w, `<LocationConstraint><LocationConstraint xmlns="http://s3.amazonaws.com/doc/2006-03-01/">`+
 370                                 h.Cluster.ClusterID+
 371                                 `</LocationConstraint></LocationConstraint>`)
 372                 } else if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 373                         // GetBucketWebsite ("GET /bucketid/?website"), GetBucketTagging, etc.
 374                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 375                 } else {
 376                         // ListObjects
 377                         h.s3list(bucketName, w, r, fs)
 378                 }
 379                 return true
 380         case r.Method == http.MethodGet || r.Method == http.MethodHead:
 381                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 382                         // GetObjectRetention ("GET /bucketid/objectid?retention&versionID=..."), etc.
 383                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 384                         return true
 385                 }
 386                 fi, err := fs.Stat(fspath)
 387                 if r.Method == "HEAD" && !objectNameGiven {
 388                         // HeadBucket
 389                         if err == nil && fi.IsDir() {
 390                                 err = setFileInfoHeaders(w.Header(), fs, fspath)
 391                                 if err != nil {
 392                                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 393                                         return true
 394                                 }
 395                                 w.WriteHeader(http.StatusOK)
 396                         } else if os.IsNotExist(err) {
 397                                 s3ErrorResponse(w, NoSuchBucket, "The specified bucket does not exist.", r.URL.Path, http.StatusNotFound)
 398                         } else {
 399                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 400                         }
 401                         return true
 402                 }
 403                 if err == nil && fi.IsDir() && objectNameGiven && strings.HasSuffix(fspath, "/") && h.Cluster.Collections.S3FolderObjects {
 404                         err = setFileInfoHeaders(w.Header(), fs, fspath)
 405                         if err != nil {
 406                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 407                                 return true
 408                         }
 409                         w.Header().Set("Content-Type", "application/x-directory")
 410                         w.WriteHeader(http.StatusOK)
 411                         return true
 412                 }
 413                 if os.IsNotExist(err) ||
 414                         (err != nil && err.Error() == "not a directory") ||
 415                         (fi != nil && fi.IsDir()) {
 416                         s3ErrorResponse(w, NoSuchKey, "The specified key does not exist.", r.URL.Path, http.StatusNotFound)
 417                         return true
 418                 }
 419
 420                 tokenUser, err := h.Cache.GetTokenUser(token)
 421                 if !h.userPermittedToUploadOrDownload(r.Method, tokenUser) {
 422                         http.Error(w, "Not permitted", http.StatusForbidden)
 423                         return true
 424                 }
 425                 h.logUploadOrDownload(r, arvclient, fs, fspath, nil, tokenUser)
 426
 427                 // shallow copy r, and change URL path
 428                 r := *r
 429                 r.URL.Path = fspath
 430                 err = setFileInfoHeaders(w.Header(), fs, fspath)
 431                 if err != nil {
 432                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 433                         return true
 434                 }
 435                 http.FileServer(fs).ServeHTTP(w, &r)
 436                 return true
 437         case r.Method == http.MethodPut:
 438                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 439                         // PutObjectAcl ("PUT /bucketid/objectid?acl&versionID=..."), etc.
 440                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 441                         return true
 442                 }
 443                 if !objectNameGiven {
 444                         s3ErrorResponse(w, InvalidArgument, "Missing object name in PUT request.", r.URL.Path, http.StatusBadRequest)
 445                         return true
 446                 }
 447                 var objectIsDir bool
 448                 if strings.HasSuffix(fspath, "/") {
 449                         if !h.Cluster.Collections.S3FolderObjects {
 450                                 s3ErrorResponse(w, InvalidArgument, "invalid object name: trailing slash", r.URL.Path, http.StatusBadRequest)
 451                                 return true
 452                         }
 453                         n, err := r.Body.Read(make([]byte, 1))
 454                         if err != nil && err != io.EOF {
 455                                 s3ErrorResponse(w, InternalError, fmt.Sprintf("error reading request body: %s", err), r.URL.Path, http.StatusInternalServerError)
 456                                 return true
 457                         } else if n > 0 {
 458                                 s3ErrorResponse(w, InvalidArgument, "cannot create object with trailing '/' char unless content is empty", r.URL.Path, http.StatusBadRequest)
 459                                 return true
 460                         } else if strings.SplitN(r.Header.Get("Content-Type"), ";", 2)[0] != "application/x-directory" {
 461                                 s3ErrorResponse(w, InvalidArgument, "cannot create object with trailing '/' char unless Content-Type is 'application/x-directory'", r.URL.Path, http.StatusBadRequest)
 462                                 return true
 463                         }
 464                         // Given PUT "foo/bar/", we'll use "foo/bar/."
 465                         // in the "ensure parents exist" block below,
 466                         // and then we'll be done.
 467                         fspath += "."
 468                         objectIsDir = true
 469                 }
 470                 fi, err := fs.Stat(fspath)
 471                 if err != nil && err.Error() == "not a directory" {
 472                         // requested foo/bar, but foo is a file
 473                         s3ErrorResponse(w, InvalidArgument, "object name conflicts with existing object", r.URL.Path, http.StatusBadRequest)
 474                         return true
 475                 }
 476                 if strings.HasSuffix(r.URL.Path, "/") && err == nil && !fi.IsDir() {
 477                         // requested foo/bar/, but foo/bar is a file
 478                         s3ErrorResponse(w, InvalidArgument, "object name conflicts with existing object", r.URL.Path, http.StatusBadRequest)
 479                         return true
 480                 }
 481                 // create missing parent/intermediate directories, if any
 482                 for i, c := range fspath {
 483                         if i > 0 && c == '/' {
 484                                 dir := fspath[:i]
 485                                 if strings.HasSuffix(dir, "/") {
 486                                         err = errors.New("invalid object name (consecutive '/' chars)")
 487                                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 488                                         return true
 489                                 }
 490                                 err = fs.Mkdir(dir, 0755)
 491                                 if errors.Is(err, arvados.ErrInvalidArgument) || errors.Is(err, arvados.ErrInvalidOperation) {
 492                                         // Cannot create a directory
 493                                         // here.
 494                                         err = fmt.Errorf("mkdir %q failed: %w", dir, err)
 495                                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 496                                         return true
 497                                 } else if err != nil && !os.IsExist(err) {
 498                                         err = fmt.Errorf("mkdir %q failed: %w", dir, err)
 499                                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 500                                         return true
 501                                 }
 502                         }
 503                 }
 504                 if !objectIsDir {
 505                         f, err := fs.OpenFile(fspath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0644)
 506                         if os.IsNotExist(err) {
 507                                 f, err = fs.OpenFile(fspath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0644)
 508                         }
 509                         if err != nil {
 510                                 err = fmt.Errorf("open %q failed: %w", r.URL.Path, err)
 511                                 s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 512                                 return true
 513                         }
 514                         defer f.Close()
 515
 516                         tokenUser, err := h.Cache.GetTokenUser(token)
 517                         if !h.userPermittedToUploadOrDownload(r.Method, tokenUser) {
 518                                 http.Error(w, "Not permitted", http.StatusForbidden)
 519                                 return true
 520                         }
 521                         h.logUploadOrDownload(r, arvclient, fs, fspath, nil, tokenUser)
 522
 523                         _, err = io.Copy(f, r.Body)
 524                         if err != nil {
 525                                 err = fmt.Errorf("write to %q failed: %w", r.URL.Path, err)
 526                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 527                                 return true
 528                         }
 529                         err = f.Close()
 530                         if err != nil {
 531                                 err = fmt.Errorf("write to %q failed: close: %w", r.URL.Path, err)
 532                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 533                                 return true
 534                         }
 535                 }
 536                 err = fs.Sync()
 537                 if err != nil {
 538                         err = fmt.Errorf("sync failed: %w", err)
 539                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 540                         return true
 541                 }
 542                 // Ensure a subsequent read operation will see the changes.
 543                 h.Cache.ResetSession(token)
 544                 w.WriteHeader(http.StatusOK)
 545                 return true
 546         case r.Method == http.MethodDelete:
 547                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 548                         // DeleteObjectTagging ("DELETE /bucketid/objectid?tagging&versionID=..."), etc.
 549                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 550                         return true
 551                 }
 552                 if !objectNameGiven || r.URL.Path == "/" {
 553                         s3ErrorResponse(w, InvalidArgument, "missing object name in DELETE request", r.URL.Path, http.StatusBadRequest)
 554                         return true
 555                 }
 556                 if strings.HasSuffix(fspath, "/") {
 557                         fspath = strings.TrimSuffix(fspath, "/")
 558                         fi, err := fs.Stat(fspath)
 559                         if os.IsNotExist(err) {
 560                                 w.WriteHeader(http.StatusNoContent)
 561                                 return true
 562                         } else if err != nil {
 563                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 564                                 return true
 565                         } else if !fi.IsDir() {
 566                                 // if "foo" exists and is a file, then
 567                                 // "foo/" doesn't exist, so we say
 568                                 // delete was successful.
 569                                 w.WriteHeader(http.StatusNoContent)
 570                                 return true
 571                         }
 572                 } else if fi, err := fs.Stat(fspath); err == nil && fi.IsDir() {
 573                         // if "foo" is a dir, it is visible via S3
 574                         // only as "foo/", not "foo" -- so we leave
 575                         // the dir alone and return 204 to indicate
 576                         // that "foo" does not exist.
 577                         w.WriteHeader(http.StatusNoContent)
 578                         return true
 579                 }
 580                 err = fs.Remove(fspath)
 581                 if os.IsNotExist(err) {
 582                         w.WriteHeader(http.StatusNoContent)
 583                         return true
 584                 }
 585                 if err != nil {
 586                         err = fmt.Errorf("rm failed: %w", err)
 587                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 588                         return true
 589                 }
 590                 err = fs.Sync()
 591                 if err != nil {
 592                         err = fmt.Errorf("sync failed: %w", err)
 593                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 594                         return true
 595                 }
 596                 // Ensure a subsequent read operation will see the changes.
 597                 h.Cache.ResetSession(token)
 598                 w.WriteHeader(http.StatusNoContent)
 599                 return true
 600         default:
 601                 s3ErrorResponse(w, InvalidRequest, "method not allowed", r.URL.Path, http.StatusMethodNotAllowed)
 602                 return true
 603         }
 604 }
 605
 606 func setFileInfoHeaders(header http.Header, fs arvados.CustomFileSystem, path string) error {
 607         path = strings.TrimSuffix(path, "/")
 608         var props map[string]interface{}
 609         for {
 610                 fi, err := fs.Stat(path)
 611                 if err != nil {
 612                         return err
 613                 }
 614                 switch src := fi.Sys().(type) {
 615                 case *arvados.Collection:
 616                         props = src.Properties
 617                 case *arvados.Group:
 618                         props = src.Properties
 619                 default:
 620                         if err, ok := src.(error); ok {
 621                                 return err
 622                         }
 623                         // Try parent
 624                         cut := strings.LastIndexByte(path, '/')
 625                         if cut < 0 {
 626                                 return nil
 627                         }
 628                         path = path[:cut]
 629                         continue
 630                 }
 631                 break
 632         }
 633         for k, v := range props {
 634                 if !validMIMEHeaderKey(k) {
 635                         continue
 636                 }
 637                 k = "x-amz-meta-" + k
 638                 if s, ok := v.(string); ok {
 639                         header.Set(k, s)
 640                 } else if j, err := json.Marshal(v); err == nil {
 641                         header.Set(k, string(j))
 642                 }
 643         }
 644         return nil
 645 }
 646
 647 func validMIMEHeaderKey(k string) bool {
 648         check := "z-" + k
 649         return check != textproto.CanonicalMIMEHeaderKey(check)
 650 }
 651
 652 // Call fn on the given path (directory) and its contents, in
 653 // lexicographic order.
 654 //
 655 // If isRoot==true and path is not a directory, return nil.
 656 //
 657 // If fn returns filepath.SkipDir when called on a directory, don't
 658 // descend into that directory.
 659 func walkFS(fs arvados.CustomFileSystem, path string, isRoot bool, fn func(path string, fi os.FileInfo) error) error {
 660         if isRoot {
 661                 fi, err := fs.Stat(path)
 662                 if os.IsNotExist(err) || (err == nil && !fi.IsDir()) {
 663                         return nil
 664                 } else if err != nil {
 665                         return err
 666                 }
 667                 err = fn(path, fi)
 668                 if err == filepath.SkipDir {
 669                         return nil
 670                 } else if err != nil {
 671                         return err
 672                 }
 673         }
 674         f, err := fs.Open(path)
 675         if os.IsNotExist(err) && isRoot {
 676                 return nil
 677         } else if err != nil {
 678                 return fmt.Errorf("open %q: %w", path, err)
 679         }
 680         defer f.Close()
 681         if path == "/" {
 682                 path = ""
 683         }
 684         fis, err := f.Readdir(-1)
 685         if err != nil {
 686                 return err
 687         }
 688         sort.Slice(fis, func(i, j int) bool { return fis[i].Name() < fis[j].Name() })
 689         for _, fi := range fis {
 690                 err = fn(path+"/"+fi.Name(), fi)
 691                 if err == filepath.SkipDir {
 692                         continue
 693                 } else if err != nil {
 694                         return err
 695                 }
 696                 if fi.IsDir() {
 697                         err = walkFS(fs, path+"/"+fi.Name(), false, fn)
 698                         if err != nil {
 699                                 return err
 700                         }
 701                 }
 702         }
 703         return nil
 704 }
 705
 706 var errDone = errors.New("done")
 707
 708 func (h *handler) s3list(bucket string, w http.ResponseWriter, r *http.Request, fs arvados.CustomFileSystem) {
 709         var params struct {
 710                 v2                bool
 711                 delimiter         string
 712                 maxKeys           int
 713                 prefix            string
 714                 marker            string // decoded continuationToken (v2) or provided by client (v1)
 715                 startAfter        string // v2
 716                 continuationToken string // v2
 717                 encodingTypeURL   bool   // v2
 718         }
 719         params.delimiter = r.FormValue("delimiter")
 720         if mk, _ := strconv.ParseInt(r.FormValue("max-keys"), 10, 64); mk > 0 && mk < s3MaxKeys {
 721                 params.maxKeys = int(mk)
 722         } else {
 723                 params.maxKeys = s3MaxKeys
 724         }
 725         params.prefix = r.FormValue("prefix")
 726         switch r.FormValue("list-type") {
 727         case "":
 728         case "2":
 729                 params.v2 = true
 730         default:
 731                 http.Error(w, "invalid list-type parameter", http.StatusBadRequest)
 732                 return
 733         }
 734         if params.v2 {
 735                 params.continuationToken = r.FormValue("continuation-token")
 736                 marker, err := base64.StdEncoding.DecodeString(params.continuationToken)
 737                 if err != nil {
 738                         http.Error(w, "invalid continuation token", http.StatusBadRequest)
 739                         return
 740                 }
 741                 params.marker = string(marker)
 742                 params.startAfter = r.FormValue("start-after")
 743                 switch r.FormValue("encoding-type") {
 744                 case "":
 745                 case "url":
 746                         params.encodingTypeURL = true
 747                 default:
 748                         http.Error(w, "invalid encoding-type parameter", http.StatusBadRequest)
 749                         return
 750                 }
 751         } else {
 752                 params.marker = r.FormValue("marker")
 753         }
 754
 755         bucketdir := "by_id/" + bucket
 756         // walkpath is the directory (relative to bucketdir) we need
 757         // to walk: the innermost directory that is guaranteed to
 758         // contain all paths that have the requested prefix. Examples:
 759         // prefix "foo/bar"  => walkpath "foo"
 760         // prefix "foo/bar/" => walkpath "foo/bar"
 761         // prefix "foo"      => walkpath ""
 762         // prefix ""         => walkpath ""
 763         walkpath := params.prefix
 764         if cut := strings.LastIndex(walkpath, "/"); cut >= 0 {
 765                 walkpath = walkpath[:cut]
 766         } else {
 767                 walkpath = ""
 768         }
 769
 770         resp := listV2Resp{
 771                 Name:              bucket,
 772                 Prefix:            params.prefix,
 773                 Delimiter:         params.delimiter,
 774                 MaxKeys:           params.maxKeys,
 775                 ContinuationToken: r.FormValue("continuation-token"),
 776                 StartAfter:        params.startAfter,
 777         }
 778         nextMarker := ""
 779
 780         commonPrefixes := map[string]bool{}
 781         err := walkFS(fs, strings.TrimSuffix(bucketdir+"/"+walkpath, "/"), true, func(path string, fi os.FileInfo) error {
 782                 if path == bucketdir {
 783                         return nil
 784                 }
 785                 path = path[len(bucketdir)+1:]
 786                 filesize := fi.Size()
 787                 if fi.IsDir() {
 788                         path += "/"
 789                         filesize = 0
 790                 }
 791                 if len(path) <= len(params.prefix) {
 792                         if path > params.prefix[:len(path)] {
 793                                 // with prefix "foobar", walking "fooz" means we're done
 794                                 return errDone
 795                         }
 796                         if path < params.prefix[:len(path)] {
 797                                 // with prefix "foobar", walking "foobag" is pointless
 798                                 return filepath.SkipDir
 799                         }
 800                         if fi.IsDir() && !strings.HasPrefix(params.prefix+"/", path) {
 801                                 // with prefix "foo/bar", walking "fo"
 802                                 // is pointless (but walking "foo" or
 803                                 // "foo/bar" is necessary)
 804                                 return filepath.SkipDir
 805                         }
 806                         if len(path) < len(params.prefix) {
 807                                 // can't skip anything, and this entry
 808                                 // isn't in the results, so just
 809                                 // continue descent
 810                                 return nil
 811                         }
 812                 } else {
 813                         if path[:len(params.prefix)] > params.prefix {
 814                                 // with prefix "foobar", nothing we
 815                                 // see after "foozzz" is relevant
 816                                 return errDone
 817                         }
 818                 }
 819                 if path < params.marker || path < params.prefix || path <= params.startAfter {
 820                         return nil
 821                 }
 822                 if fi.IsDir() && !h.Cluster.Collections.S3FolderObjects {
 823                         // Note we don't add anything to
 824                         // commonPrefixes here even if delimiter is
 825                         // "/". We descend into the directory, and
 826                         // return a commonPrefix only if we end up
 827                         // finding a regular file inside it.
 828                         return nil
 829                 }
 830                 if len(resp.Contents)+len(commonPrefixes) >= params.maxKeys {
 831                         resp.IsTruncated = true
 832                         if params.delimiter != "" || params.v2 {
 833                                 nextMarker = path
 834                         }
 835                         return errDone
 836                 }
 837                 if params.delimiter != "" {
 838                         idx := strings.Index(path[len(params.prefix):], params.delimiter)
 839                         if idx >= 0 {
 840                                 // with prefix "foobar" and delimiter
 841                                 // "z", when we hit "foobar/baz", we
 842                                 // add "/baz" to commonPrefixes and
 843                                 // stop descending.
 844                                 commonPrefixes[path[:len(params.prefix)+idx+1]] = true
 845                                 return filepath.SkipDir
 846                         }
 847                 }
 848                 resp.Contents = append(resp.Contents, s3.Key{
 849                         Key:          path,
 850                         LastModified: fi.ModTime().UTC().Format("2006-01-02T15:04:05.999") + "Z",
 851                         Size:         filesize,
 852                 })
 853                 return nil
 854         })
 855         if err != nil && err != errDone {
 856                 http.Error(w, err.Error(), http.StatusInternalServerError)
 857                 return
 858         }
 859         if params.delimiter != "" {
 860                 resp.CommonPrefixes = make([]commonPrefix, 0, len(commonPrefixes))
 861                 for prefix := range commonPrefixes {
 862                         resp.CommonPrefixes = append(resp.CommonPrefixes, commonPrefix{prefix})
 863                 }
 864                 sort.Slice(resp.CommonPrefixes, func(i, j int) bool { return resp.CommonPrefixes[i].Prefix < resp.CommonPrefixes[j].Prefix })
 865         }
 866         resp.KeyCount = len(resp.Contents)
 867         var respV1orV2 interface{}
 868
 869         if params.encodingTypeURL {
 870                 // https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html
 871                 // "If you specify the encoding-type request
 872                 // parameter, Amazon S3 includes this element in the
 873                 // response, and returns encoded key name values in
 874                 // the following response elements:
 875                 //
 876                 // Delimiter, Prefix, Key, and StartAfter.
 877                 //
 878                 //      Type: String
 879                 //
 880                 // Valid Values: url"
 881                 //
 882                 // This is somewhat vague but in practice it appears
 883                 // to mean x-www-form-urlencoded as in RFC1866 8.2.1
 884                 // para 1 (encode space as "+") rather than straight
 885                 // percent-encoding as in RFC1738 2.2.  Presumably,
 886                 // the intent is to allow the client to decode XML and
 887                 // then paste the strings directly into another URI
 888                 // query or POST form like "https://host/path?foo=" +
 889                 // foo + "&bar=" + bar.
 890                 resp.EncodingType = "url"
 891                 resp.Delimiter = url.QueryEscape(resp.Delimiter)
 892                 resp.Prefix = url.QueryEscape(resp.Prefix)
 893                 resp.StartAfter = url.QueryEscape(resp.StartAfter)
 894                 for i, ent := range resp.Contents {
 895                         ent.Key = url.QueryEscape(ent.Key)
 896                         resp.Contents[i] = ent
 897                 }
 898                 for i, ent := range resp.CommonPrefixes {
 899                         ent.Prefix = url.QueryEscape(ent.Prefix)
 900                         resp.CommonPrefixes[i] = ent
 901                 }
 902         }
 903
 904         if params.v2 {
 905                 resp.NextContinuationToken = base64.StdEncoding.EncodeToString([]byte(nextMarker))
 906                 respV1orV2 = resp
 907         } else {
 908                 respV1orV2 = listV1Resp{
 909                         CommonPrefixes: resp.CommonPrefixes,
 910                         NextMarker:     nextMarker,
 911                         KeyCount:       resp.KeyCount,
 912                         ListResp: s3.ListResp{
 913                                 IsTruncated: resp.IsTruncated,
 914                                 Name:        bucket,
 915                                 Prefix:      params.prefix,
 916                                 Delimiter:   params.delimiter,
 917                                 Marker:      params.marker,
 918                                 MaxKeys:     params.maxKeys,
 919                                 Contents:    resp.Contents,
 920                         },
 921                 }
 922         }
 923
 924         w.Header().Set("Content-Type", "application/xml")
 925         io.WriteString(w, xml.Header)
 926         if err := xml.NewEncoder(w).Encode(respV1orV2); err != nil {
 927                 ctxlog.FromContext(r.Context()).WithError(err).Error("error writing xml response")
 928         }
 929 }