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                                 setFileInfoHeaders(w.Header(), fs, fspath)
 391                                 w.WriteHeader(http.StatusOK)
 392                         } else if os.IsNotExist(err) {
 393                                 s3ErrorResponse(w, NoSuchBucket, "The specified bucket does not exist.", r.URL.Path, http.StatusNotFound)
 394                         } else {
 395                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 396                         }
 397                         return true
 398                 }
 399                 if err == nil && fi.IsDir() && objectNameGiven && strings.HasSuffix(fspath, "/") && h.Cluster.Collections.S3FolderObjects {
 400                         setFileInfoHeaders(w.Header(), fs, fspath)
 401                         w.Header().Set("Content-Type", "application/x-directory")
 402                         w.WriteHeader(http.StatusOK)
 403                         return true
 404                 }
 405                 if os.IsNotExist(err) ||
 406                         (err != nil && err.Error() == "not a directory") ||
 407                         (fi != nil && fi.IsDir()) {
 408                         s3ErrorResponse(w, NoSuchKey, "The specified key does not exist.", r.URL.Path, http.StatusNotFound)
 409                         return true
 410                 }
 411
 412                 tokenUser, err := h.Cache.GetTokenUser(token)
 413                 if !h.userPermittedToUploadOrDownload(r.Method, tokenUser) {
 414                         http.Error(w, "Not permitted", http.StatusForbidden)
 415                         return true
 416                 }
 417                 h.logUploadOrDownload(r, arvclient, fs, fspath, nil, tokenUser)
 418
 419                 // shallow copy r, and change URL path
 420                 r := *r
 421                 r.URL.Path = fspath
 422                 setFileInfoHeaders(w.Header(), fs, fspath)
 423                 http.FileServer(fs).ServeHTTP(w, &r)
 424                 return true
 425         case r.Method == http.MethodPut:
 426                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 427                         // PutObjectAcl ("PUT /bucketid/objectid?acl&versionID=..."), etc.
 428                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 429                         return true
 430                 }
 431                 if !objectNameGiven {
 432                         s3ErrorResponse(w, InvalidArgument, "Missing object name in PUT request.", r.URL.Path, http.StatusBadRequest)
 433                         return true
 434                 }
 435                 var objectIsDir bool
 436                 if strings.HasSuffix(fspath, "/") {
 437                         if !h.Cluster.Collections.S3FolderObjects {
 438                                 s3ErrorResponse(w, InvalidArgument, "invalid object name: trailing slash", r.URL.Path, http.StatusBadRequest)
 439                                 return true
 440                         }
 441                         n, err := r.Body.Read(make([]byte, 1))
 442                         if err != nil && err != io.EOF {
 443                                 s3ErrorResponse(w, InternalError, fmt.Sprintf("error reading request body: %s", err), r.URL.Path, http.StatusInternalServerError)
 444                                 return true
 445                         } else if n > 0 {
 446                                 s3ErrorResponse(w, InvalidArgument, "cannot create object with trailing '/' char unless content is empty", r.URL.Path, http.StatusBadRequest)
 447                                 return true
 448                         } else if strings.SplitN(r.Header.Get("Content-Type"), ";", 2)[0] != "application/x-directory" {
 449                                 s3ErrorResponse(w, InvalidArgument, "cannot create object with trailing '/' char unless Content-Type is 'application/x-directory'", r.URL.Path, http.StatusBadRequest)
 450                                 return true
 451                         }
 452                         // Given PUT "foo/bar/", we'll use "foo/bar/."
 453                         // in the "ensure parents exist" block below,
 454                         // and then we'll be done.
 455                         fspath += "."
 456                         objectIsDir = true
 457                 }
 458                 fi, err := fs.Stat(fspath)
 459                 if err != nil && err.Error() == "not a directory" {
 460                         // requested foo/bar, but foo is a file
 461                         s3ErrorResponse(w, InvalidArgument, "object name conflicts with existing object", r.URL.Path, http.StatusBadRequest)
 462                         return true
 463                 }
 464                 if strings.HasSuffix(r.URL.Path, "/") && err == nil && !fi.IsDir() {
 465                         // requested foo/bar/, but foo/bar is a file
 466                         s3ErrorResponse(w, InvalidArgument, "object name conflicts with existing object", r.URL.Path, http.StatusBadRequest)
 467                         return true
 468                 }
 469                 // create missing parent/intermediate directories, if any
 470                 for i, c := range fspath {
 471                         if i > 0 && c == '/' {
 472                                 dir := fspath[:i]
 473                                 if strings.HasSuffix(dir, "/") {
 474                                         err = errors.New("invalid object name (consecutive '/' chars)")
 475                                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 476                                         return true
 477                                 }
 478                                 err = fs.Mkdir(dir, 0755)
 479                                 if errors.Is(err, arvados.ErrInvalidArgument) || errors.Is(err, arvados.ErrInvalidOperation) {
 480                                         // Cannot create a directory
 481                                         // here.
 482                                         err = fmt.Errorf("mkdir %q failed: %w", dir, err)
 483                                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 484                                         return true
 485                                 } else if err != nil && !os.IsExist(err) {
 486                                         err = fmt.Errorf("mkdir %q failed: %w", dir, err)
 487                                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 488                                         return true
 489                                 }
 490                         }
 491                 }
 492                 if !objectIsDir {
 493                         f, err := fs.OpenFile(fspath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0644)
 494                         if os.IsNotExist(err) {
 495                                 f, err = fs.OpenFile(fspath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0644)
 496                         }
 497                         if err != nil {
 498                                 err = fmt.Errorf("open %q failed: %w", r.URL.Path, err)
 499                                 s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 500                                 return true
 501                         }
 502                         defer f.Close()
 503
 504                         tokenUser, err := h.Cache.GetTokenUser(token)
 505                         if !h.userPermittedToUploadOrDownload(r.Method, tokenUser) {
 506                                 http.Error(w, "Not permitted", http.StatusForbidden)
 507                                 return true
 508                         }
 509                         h.logUploadOrDownload(r, arvclient, fs, fspath, nil, tokenUser)
 510
 511                         _, err = io.Copy(f, r.Body)
 512                         if err != nil {
 513                                 err = fmt.Errorf("write to %q failed: %w", r.URL.Path, err)
 514                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 515                                 return true
 516                         }
 517                         err = f.Close()
 518                         if err != nil {
 519                                 err = fmt.Errorf("write to %q failed: close: %w", r.URL.Path, err)
 520                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 521                                 return true
 522                         }
 523                 }
 524                 err = fs.Sync()
 525                 if err != nil {
 526                         err = fmt.Errorf("sync failed: %w", err)
 527                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 528                         return true
 529                 }
 530                 // Ensure a subsequent read operation will see the changes.
 531                 h.Cache.ResetSession(token)
 532                 w.WriteHeader(http.StatusOK)
 533                 return true
 534         case r.Method == http.MethodDelete:
 535                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 536                         // DeleteObjectTagging ("DELETE /bucketid/objectid?tagging&versionID=..."), etc.
 537                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 538                         return true
 539                 }
 540                 if !objectNameGiven || r.URL.Path == "/" {
 541                         s3ErrorResponse(w, InvalidArgument, "missing object name in DELETE request", r.URL.Path, http.StatusBadRequest)
 542                         return true
 543                 }
 544                 if strings.HasSuffix(fspath, "/") {
 545                         fspath = strings.TrimSuffix(fspath, "/")
 546                         fi, err := fs.Stat(fspath)
 547                         if os.IsNotExist(err) {
 548                                 w.WriteHeader(http.StatusNoContent)
 549                                 return true
 550                         } else if err != nil {
 551                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 552                                 return true
 553                         } else if !fi.IsDir() {
 554                                 // if "foo" exists and is a file, then
 555                                 // "foo/" doesn't exist, so we say
 556                                 // delete was successful.
 557                                 w.WriteHeader(http.StatusNoContent)
 558                                 return true
 559                         }
 560                 } else if fi, err := fs.Stat(fspath); err == nil && fi.IsDir() {
 561                         // if "foo" is a dir, it is visible via S3
 562                         // only as "foo/", not "foo" -- so we leave
 563                         // the dir alone and return 204 to indicate
 564                         // that "foo" does not exist.
 565                         w.WriteHeader(http.StatusNoContent)
 566                         return true
 567                 }
 568                 err = fs.Remove(fspath)
 569                 if os.IsNotExist(err) {
 570                         w.WriteHeader(http.StatusNoContent)
 571                         return true
 572                 }
 573                 if err != nil {
 574                         err = fmt.Errorf("rm failed: %w", err)
 575                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 576                         return true
 577                 }
 578                 err = fs.Sync()
 579                 if err != nil {
 580                         err = fmt.Errorf("sync failed: %w", err)
 581                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 582                         return true
 583                 }
 584                 // Ensure a subsequent read operation will see the changes.
 585                 h.Cache.ResetSession(token)
 586                 w.WriteHeader(http.StatusNoContent)
 587                 return true
 588         default:
 589                 s3ErrorResponse(w, InvalidRequest, "method not allowed", r.URL.Path, http.StatusMethodNotAllowed)
 590                 return true
 591         }
 592 }
 593
 594 func setFileInfoHeaders(header http.Header, fs arvados.CustomFileSystem, path string) {
 595         path = strings.TrimSuffix(path, "/")
 596         var props map[string]interface{}
 597         for {
 598                 fi, err := fs.Stat(path)
 599                 if err != nil {
 600                         return
 601                 }
 602                 switch src := fi.Sys().(type) {
 603                 case *arvados.Collection:
 604                         props = src.Properties
 605                 case *arvados.Group:
 606                         props = src.Properties
 607                 default:
 608                         // Try parent
 609                         cut := strings.LastIndexByte(path, '/')
 610                         if cut < 0 {
 611                                 return
 612                         }
 613                         path = path[:cut]
 614                         continue
 615                 }
 616                 break
 617         }
 618         for k, v := range props {
 619                 if !validMIMEHeaderKey(k) {
 620                         continue
 621                 }
 622                 k = "x-amz-meta-" + k
 623                 if s, ok := v.(string); ok {
 624                         header.Set(k, s)
 625                 } else if j, err := json.Marshal(v); err == nil {
 626                         header.Set(k, string(j))
 627                 }
 628         }
 629 }
 630
 631 func validMIMEHeaderKey(k string) bool {
 632         check := "z-" + k
 633         return check != textproto.CanonicalMIMEHeaderKey(check)
 634 }
 635
 636 // Call fn on the given path (directory) and its contents, in
 637 // lexicographic order.
 638 //
 639 // If isRoot==true and path is not a directory, return nil.
 640 //
 641 // If fn returns filepath.SkipDir when called on a directory, don't
 642 // descend into that directory.
 643 func walkFS(fs arvados.CustomFileSystem, path string, isRoot bool, fn func(path string, fi os.FileInfo) error) error {
 644         if isRoot {
 645                 fi, err := fs.Stat(path)
 646                 if os.IsNotExist(err) || (err == nil && !fi.IsDir()) {
 647                         return nil
 648                 } else if err != nil {
 649                         return err
 650                 }
 651                 err = fn(path, fi)
 652                 if err == filepath.SkipDir {
 653                         return nil
 654                 } else if err != nil {
 655                         return err
 656                 }
 657         }
 658         f, err := fs.Open(path)
 659         if os.IsNotExist(err) && isRoot {
 660                 return nil
 661         } else if err != nil {
 662                 return fmt.Errorf("open %q: %w", path, err)
 663         }
 664         defer f.Close()
 665         if path == "/" {
 666                 path = ""
 667         }
 668         fis, err := f.Readdir(-1)
 669         if err != nil {
 670                 return err
 671         }
 672         sort.Slice(fis, func(i, j int) bool { return fis[i].Name() < fis[j].Name() })
 673         for _, fi := range fis {
 674                 err = fn(path+"/"+fi.Name(), fi)
 675                 if err == filepath.SkipDir {
 676                         continue
 677                 } else if err != nil {
 678                         return err
 679                 }
 680                 if fi.IsDir() {
 681                         err = walkFS(fs, path+"/"+fi.Name(), false, fn)
 682                         if err != nil {
 683                                 return err
 684                         }
 685                 }
 686         }
 687         return nil
 688 }
 689
 690 var errDone = errors.New("done")
 691
 692 func (h *handler) s3list(bucket string, w http.ResponseWriter, r *http.Request, fs arvados.CustomFileSystem) {
 693         var params struct {
 694                 v2                bool
 695                 delimiter         string
 696                 maxKeys           int
 697                 prefix            string
 698                 marker            string // decoded continuationToken (v2) or provided by client (v1)
 699                 startAfter        string // v2
 700                 continuationToken string // v2
 701                 encodingTypeURL   bool   // v2
 702         }
 703         params.delimiter = r.FormValue("delimiter")
 704         if mk, _ := strconv.ParseInt(r.FormValue("max-keys"), 10, 64); mk > 0 && mk < s3MaxKeys {
 705                 params.maxKeys = int(mk)
 706         } else {
 707                 params.maxKeys = s3MaxKeys
 708         }
 709         params.prefix = r.FormValue("prefix")
 710         switch r.FormValue("list-type") {
 711         case "":
 712         case "2":
 713                 params.v2 = true
 714         default:
 715                 http.Error(w, "invalid list-type parameter", http.StatusBadRequest)
 716                 return
 717         }
 718         if params.v2 {
 719                 params.continuationToken = r.FormValue("continuation-token")
 720                 marker, err := base64.StdEncoding.DecodeString(params.continuationToken)
 721                 if err != nil {
 722                         http.Error(w, "invalid continuation token", http.StatusBadRequest)
 723                         return
 724                 }
 725                 params.marker = string(marker)
 726                 params.startAfter = r.FormValue("start-after")
 727                 switch r.FormValue("encoding-type") {
 728                 case "":
 729                 case "url":
 730                         params.encodingTypeURL = true
 731                 default:
 732                         http.Error(w, "invalid encoding-type parameter", http.StatusBadRequest)
 733                         return
 734                 }
 735         } else {
 736                 params.marker = r.FormValue("marker")
 737         }
 738
 739         bucketdir := "by_id/" + bucket
 740         // walkpath is the directory (relative to bucketdir) we need
 741         // to walk: the innermost directory that is guaranteed to
 742         // contain all paths that have the requested prefix. Examples:
 743         // prefix "foo/bar"  => walkpath "foo"
 744         // prefix "foo/bar/" => walkpath "foo/bar"
 745         // prefix "foo"      => walkpath ""
 746         // prefix ""         => walkpath ""
 747         walkpath := params.prefix
 748         if cut := strings.LastIndex(walkpath, "/"); cut >= 0 {
 749                 walkpath = walkpath[:cut]
 750         } else {
 751                 walkpath = ""
 752         }
 753
 754         resp := listV2Resp{
 755                 Name:              bucket,
 756                 Prefix:            params.prefix,
 757                 Delimiter:         params.delimiter,
 758                 MaxKeys:           params.maxKeys,
 759                 ContinuationToken: r.FormValue("continuation-token"),
 760                 StartAfter:        params.startAfter,
 761         }
 762         nextMarker := ""
 763
 764         commonPrefixes := map[string]bool{}
 765         err := walkFS(fs, strings.TrimSuffix(bucketdir+"/"+walkpath, "/"), true, func(path string, fi os.FileInfo) error {
 766                 if path == bucketdir {
 767                         return nil
 768                 }
 769                 path = path[len(bucketdir)+1:]
 770                 filesize := fi.Size()
 771                 if fi.IsDir() {
 772                         path += "/"
 773                         filesize = 0
 774                 }
 775                 if len(path) <= len(params.prefix) {
 776                         if path > params.prefix[:len(path)] {
 777                                 // with prefix "foobar", walking "fooz" means we're done
 778                                 return errDone
 779                         }
 780                         if path < params.prefix[:len(path)] {
 781                                 // with prefix "foobar", walking "foobag" is pointless
 782                                 return filepath.SkipDir
 783                         }
 784                         if fi.IsDir() && !strings.HasPrefix(params.prefix+"/", path) {
 785                                 // with prefix "foo/bar", walking "fo"
 786                                 // is pointless (but walking "foo" or
 787                                 // "foo/bar" is necessary)
 788                                 return filepath.SkipDir
 789                         }
 790                         if len(path) < len(params.prefix) {
 791                                 // can't skip anything, and this entry
 792                                 // isn't in the results, so just
 793                                 // continue descent
 794                                 return nil
 795                         }
 796                 } else {
 797                         if path[:len(params.prefix)] > params.prefix {
 798                                 // with prefix "foobar", nothing we
 799                                 // see after "foozzz" is relevant
 800                                 return errDone
 801                         }
 802                 }
 803                 if path < params.marker || path < params.prefix || path <= params.startAfter {
 804                         return nil
 805                 }
 806                 if fi.IsDir() && !h.Cluster.Collections.S3FolderObjects {
 807                         // Note we don't add anything to
 808                         // commonPrefixes here even if delimiter is
 809                         // "/". We descend into the directory, and
 810                         // return a commonPrefix only if we end up
 811                         // finding a regular file inside it.
 812                         return nil
 813                 }
 814                 if len(resp.Contents)+len(commonPrefixes) >= params.maxKeys {
 815                         resp.IsTruncated = true
 816                         if params.delimiter != "" || params.v2 {
 817                                 nextMarker = path
 818                         }
 819                         return errDone
 820                 }
 821                 if params.delimiter != "" {
 822                         idx := strings.Index(path[len(params.prefix):], params.delimiter)
 823                         if idx >= 0 {
 824                                 // with prefix "foobar" and delimiter
 825                                 // "z", when we hit "foobar/baz", we
 826                                 // add "/baz" to commonPrefixes and
 827                                 // stop descending.
 828                                 commonPrefixes[path[:len(params.prefix)+idx+1]] = true
 829                                 return filepath.SkipDir
 830                         }
 831                 }
 832                 resp.Contents = append(resp.Contents, s3.Key{
 833                         Key:          path,
 834                         LastModified: fi.ModTime().UTC().Format("2006-01-02T15:04:05.999") + "Z",
 835                         Size:         filesize,
 836                 })
 837                 return nil
 838         })
 839         if err != nil && err != errDone {
 840                 http.Error(w, err.Error(), http.StatusInternalServerError)
 841                 return
 842         }
 843         if params.delimiter != "" {
 844                 resp.CommonPrefixes = make([]commonPrefix, 0, len(commonPrefixes))
 845                 for prefix := range commonPrefixes {
 846                         resp.CommonPrefixes = append(resp.CommonPrefixes, commonPrefix{prefix})
 847                 }
 848                 sort.Slice(resp.CommonPrefixes, func(i, j int) bool { return resp.CommonPrefixes[i].Prefix < resp.CommonPrefixes[j].Prefix })
 849         }
 850         resp.KeyCount = len(resp.Contents)
 851         var respV1orV2 interface{}
 852
 853         if params.encodingTypeURL {
 854                 // https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html
 855                 // "If you specify the encoding-type request
 856                 // parameter, Amazon S3 includes this element in the
 857                 // response, and returns encoded key name values in
 858                 // the following response elements:
 859                 //
 860                 // Delimiter, Prefix, Key, and StartAfter.
 861                 //
 862                 //      Type: String
 863                 //
 864                 // Valid Values: url"
 865                 //
 866                 // This is somewhat vague but in practice it appears
 867                 // to mean x-www-form-urlencoded as in RFC1866 8.2.1
 868                 // para 1 (encode space as "+") rather than straight
 869                 // percent-encoding as in RFC1738 2.2.  Presumably,
 870                 // the intent is to allow the client to decode XML and
 871                 // then paste the strings directly into another URI
 872                 // query or POST form like "https://host/path?foo=" +
 873                 // foo + "&bar=" + bar.
 874                 resp.EncodingType = "url"
 875                 resp.Delimiter = url.QueryEscape(resp.Delimiter)
 876                 resp.Prefix = url.QueryEscape(resp.Prefix)
 877                 resp.StartAfter = url.QueryEscape(resp.StartAfter)
 878                 for i, ent := range resp.Contents {
 879                         ent.Key = url.QueryEscape(ent.Key)
 880                         resp.Contents[i] = ent
 881                 }
 882                 for i, ent := range resp.CommonPrefixes {
 883                         ent.Prefix = url.QueryEscape(ent.Prefix)
 884                         resp.CommonPrefixes[i] = ent
 885                 }
 886         }
 887
 888         if params.v2 {
 889                 resp.NextContinuationToken = base64.StdEncoding.EncodeToString([]byte(nextMarker))
 890                 respV1orV2 = resp
 891         } else {
 892                 respV1orV2 = listV1Resp{
 893                         CommonPrefixes: resp.CommonPrefixes,
 894                         NextMarker:     nextMarker,
 895                         KeyCount:       resp.KeyCount,
 896                         ListResp: s3.ListResp{
 897                                 IsTruncated: resp.IsTruncated,
 898                                 Name:        bucket,
 899                                 Prefix:      params.prefix,
 900                                 Delimiter:   params.delimiter,
 901                                 Marker:      params.marker,
 902                                 MaxKeys:     params.maxKeys,
 903                                 Contents:    resp.Contents,
 904                         },
 905                 }
 906         }
 907
 908         w.Header().Set("Content-Type", "application/xml")
 909         io.WriteString(w, xml.Header)
 910         if err := xml.NewEncoder(w).Encode(respV1orV2); err != nil {
 911                 ctxlog.FromContext(r.Context()).WithError(err).Error("error writing xml response")
 912         }
 913 }