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         "mime"
  18         "net/http"
  19         "net/textproto"
  20         "net/url"
  21         "os"
  22         "path/filepath"
  23         "regexp"
  24         "sort"
  25         "strconv"
  26         "strings"
  27         "time"
  28
  29         "git.arvados.org/arvados.git/sdk/go/arvados"
  30         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  31         "github.com/AdRoll/goamz/s3"
  32 )
  33
  34 const (
  35         s3MaxKeys       = 1000
  36         s3SignAlgorithm = "AWS4-HMAC-SHA256"
  37         s3MaxClockSkew  = 5 * time.Minute
  38 )
  39
  40 type commonPrefix struct {
  41         Prefix string
  42 }
  43
  44 type listV1Resp struct {
  45         XMLName string `xml:"http://s3.amazonaws.com/doc/2006-03-01/ ListBucketResult"`
  46         s3.ListResp
  47         // s3.ListResp marshals an empty tag when
  48         // CommonPrefixes is nil, which confuses some clients.
  49         // Fix by using this nested struct instead.
  50         CommonPrefixes []commonPrefix
  51         // Similarly, we need omitempty here, because an empty
  52         // tag confuses some clients (e.g.,
  53         // github.com/aws/aws-sdk-net never terminates its
  54         // paging loop).
  55         NextMarker string `xml:"NextMarker,omitempty"`
  56         // ListObjectsV2 has a KeyCount response field.
  57         KeyCount int
  58 }
  59
  60 type listV2Resp struct {
  61         XMLName               string `xml:"http://s3.amazonaws.com/doc/2006-03-01/ ListBucketResult"`
  62         IsTruncated           bool
  63         Contents              []s3.Key
  64         Name                  string
  65         Prefix                string
  66         Delimiter             string
  67         MaxKeys               int
  68         CommonPrefixes        []commonPrefix
  69         EncodingType          string `xml:",omitempty"`
  70         KeyCount              int
  71         ContinuationToken     string `xml:",omitempty"`
  72         NextContinuationToken string `xml:",omitempty"`
  73         StartAfter            string `xml:",omitempty"`
  74 }
  75
  76 func hmacstring(msg string, key []byte) []byte {
  77         h := hmac.New(sha256.New, key)
  78         io.WriteString(h, msg)
  79         return h.Sum(nil)
  80 }
  81
  82 func hashdigest(h hash.Hash, payload string) string {
  83         io.WriteString(h, payload)
  84         return fmt.Sprintf("%x", h.Sum(nil))
  85 }
  86
  87 // Signing key for given secret key and request attrs.
  88 func s3signatureKey(key, datestamp, regionName, serviceName string) []byte {
  89         return hmacstring("aws4_request",
  90                 hmacstring(serviceName,
  91                         hmacstring(regionName,
  92                                 hmacstring(datestamp, []byte("AWS4"+key)))))
  93 }
  94
  95 // Canonical query string for S3 V4 signature: sorted keys, spaces
  96 // escaped as %20 instead of +, keyvalues joined with &.
  97 func s3querystring(u *url.URL) string {
  98         keys := make([]string, 0, len(u.Query()))
  99         values := make(map[string]string, len(u.Query()))
 100         for k, vs := range u.Query() {
 101                 k = strings.Replace(url.QueryEscape(k), "+", "%20", -1)
 102                 keys = append(keys, k)
 103                 for _, v := range vs {
 104                         v = strings.Replace(url.QueryEscape(v), "+", "%20", -1)
 105                         if values[k] != "" {
 106                                 values[k] += "&"
 107                         }
 108                         values[k] += k + "=" + v
 109                 }
 110         }
 111         sort.Strings(keys)
 112         for i, k := range keys {
 113                 keys[i] = values[k]
 114         }
 115         return strings.Join(keys, "&")
 116 }
 117
 118 var reMultipleSlashChars = regexp.MustCompile(`//+`)
 119
 120 func s3stringToSign(alg, scope, signedHeaders string, r *http.Request) (string, error) {
 121         timefmt, timestr := "20060102T150405Z", r.Header.Get("X-Amz-Date")
 122         if timestr == "" {
 123                 timefmt, timestr = time.RFC1123, r.Header.Get("Date")
 124         }
 125         t, err := time.Parse(timefmt, timestr)
 126         if err != nil {
 127                 return "", fmt.Errorf("invalid timestamp %q: %s", timestr, err)
 128         }
 129         if skew := time.Now().Sub(t); skew < -s3MaxClockSkew || skew > s3MaxClockSkew {
 130                 return "", errors.New("exceeded max clock skew")
 131         }
 132
 133         var canonicalHeaders string
 134         for _, h := range strings.Split(signedHeaders, ";") {
 135                 if h == "host" {
 136                         canonicalHeaders += h + ":" + r.Host + "\n"
 137                 } else {
 138                         canonicalHeaders += h + ":" + r.Header.Get(h) + "\n"
 139                 }
 140         }
 141
 142         normalizedPath := normalizePath(r.URL.Path)
 143         ctxlog.FromContext(r.Context()).Debugf("normalizedPath %q", normalizedPath)
 144         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"))
 145         ctxlog.FromContext(r.Context()).Debugf("s3stringToSign: canonicalRequest %s", canonicalRequest)
 146         return fmt.Sprintf("%s\n%s\n%s\n%s", alg, r.Header.Get("X-Amz-Date"), scope, hashdigest(sha256.New(), canonicalRequest)), nil
 147 }
 148
 149 func normalizePath(s string) string {
 150         // (url.URL).EscapedPath() would be incorrect here. AWS
 151         // documentation specifies the URL path should be normalized
 152         // according to RFC 3986, i.e., unescaping ALPHA / DIGIT / "-"
 153         // / "." / "_" / "~". The implication is that everything other
 154         // than those chars (and "/") _must_ be percent-encoded --
 155         // even chars like ";" and "," that are not normally
 156         // percent-encoded in paths.
 157         out := ""
 158         for _, c := range []byte(reMultipleSlashChars.ReplaceAllString(s, "/")) {
 159                 if (c >= 'a' && c <= 'z') ||
 160                         (c >= 'A' && c <= 'Z') ||
 161                         (c >= '0' && c <= '9') ||
 162                         c == '-' ||
 163                         c == '.' ||
 164                         c == '_' ||
 165                         c == '~' ||
 166                         c == '/' {
 167                         out += string(c)
 168                 } else {
 169                         out += fmt.Sprintf("%%%02X", c)
 170                 }
 171         }
 172         return out
 173 }
 174
 175 func s3signature(secretKey, scope, signedHeaders, stringToSign string) (string, error) {
 176         // scope is {datestamp}/{region}/{service}/aws4_request
 177         drs := strings.Split(scope, "/")
 178         if len(drs) != 4 {
 179                 return "", fmt.Errorf("invalid scope %q", scope)
 180         }
 181         key := s3signatureKey(secretKey, drs[0], drs[1], drs[2])
 182         return hashdigest(hmac.New(sha256.New, key), stringToSign), nil
 183 }
 184
 185 var v2tokenUnderscore = regexp.MustCompile(`^v2_[a-z0-9]{5}-gj3su-[a-z0-9]{15}_`)
 186
 187 func unescapeKey(key string) string {
 188         if v2tokenUnderscore.MatchString(key) {
 189                 // Entire Arvados token, with "/" replaced by "_" to
 190                 // avoid colliding with the Authorization header
 191                 // format.
 192                 return strings.Replace(key, "_", "/", -1)
 193         } else if s, err := url.PathUnescape(key); err == nil {
 194                 return s
 195         } else {
 196                 return key
 197         }
 198 }
 199
 200 // checks3signature verifies the given S3 V4 signature and returns the
 201 // Arvados token that corresponds to the given accessKey. An error is
 202 // returned if accessKey is not a valid token UUID or the signature
 203 // does not match.
 204 func (h *handler) checks3signature(r *http.Request) (string, error) {
 205         var key, scope, signedHeaders, signature string
 206         authstring := strings.TrimPrefix(r.Header.Get("Authorization"), s3SignAlgorithm+" ")
 207         for _, cmpt := range strings.Split(authstring, ",") {
 208                 cmpt = strings.TrimSpace(cmpt)
 209                 split := strings.SplitN(cmpt, "=", 2)
 210                 switch {
 211                 case len(split) != 2:
 212                         // (?) ignore
 213                 case split[0] == "Credential":
 214                         keyandscope := strings.SplitN(split[1], "/", 2)
 215                         if len(keyandscope) == 2 {
 216                                 key, scope = keyandscope[0], keyandscope[1]
 217                         }
 218                 case split[0] == "SignedHeaders":
 219                         signedHeaders = split[1]
 220                 case split[0] == "Signature":
 221                         signature = split[1]
 222                 }
 223         }
 224
 225         client := (&arvados.Client{
 226                 APIHost:  h.Cluster.Services.Controller.ExternalURL.Host,
 227                 Insecure: h.Cluster.TLS.Insecure,
 228         }).WithRequestID(r.Header.Get("X-Request-Id"))
 229         var aca arvados.APIClientAuthorization
 230         var secret string
 231         var err error
 232         if len(key) == 27 && key[5:12] == "-gj3su-" {
 233                 // Access key is the UUID of an Arvados token, secret
 234                 // key is the secret part.
 235                 ctx := arvados.ContextWithAuthorization(r.Context(), "Bearer "+h.Cluster.SystemRootToken)
 236                 err = client.RequestAndDecodeContext(ctx, &aca, "GET", "arvados/v1/api_client_authorizations/"+key, nil, nil)
 237                 secret = aca.APIToken
 238         } else {
 239                 // Access key and secret key are both an entire
 240                 // Arvados token or OIDC access token.
 241                 ctx := arvados.ContextWithAuthorization(r.Context(), "Bearer "+unescapeKey(key))
 242                 err = client.RequestAndDecodeContext(ctx, &aca, "GET", "arvados/v1/api_client_authorizations/current", nil, nil)
 243                 secret = key
 244         }
 245         if err != nil {
 246                 ctxlog.FromContext(r.Context()).WithError(err).WithField("UUID", key).Info("token lookup failed")
 247                 return "", errors.New("invalid access key")
 248         }
 249         stringToSign, err := s3stringToSign(s3SignAlgorithm, scope, signedHeaders, r)
 250         if err != nil {
 251                 return "", err
 252         }
 253         expect, err := s3signature(secret, scope, signedHeaders, stringToSign)
 254         if err != nil {
 255                 return "", err
 256         } else if expect != signature {
 257                 return "", fmt.Errorf("signature does not match (scope %q signedHeaders %q stringToSign %q)", scope, signedHeaders, stringToSign)
 258         }
 259         return aca.TokenV2(), nil
 260 }
 261
 262 func s3ErrorResponse(w http.ResponseWriter, s3code string, message string, resource string, code int) {
 263         w.Header().Set("Content-Type", "application/xml")
 264         w.Header().Set("X-Content-Type-Options", "nosniff")
 265         w.WriteHeader(code)
 266         var errstruct struct {
 267                 Code      string
 268                 Message   string
 269                 Resource  string
 270                 RequestId string
 271         }
 272         errstruct.Code = s3code
 273         errstruct.Message = message
 274         errstruct.Resource = resource
 275         errstruct.RequestId = ""
 276         enc := xml.NewEncoder(w)
 277         fmt.Fprint(w, xml.Header)
 278         enc.EncodeElement(errstruct, xml.StartElement{Name: xml.Name{Local: "Error"}})
 279 }
 280
 281 var NoSuchKey = "NoSuchKey"
 282 var NoSuchBucket = "NoSuchBucket"
 283 var InvalidArgument = "InvalidArgument"
 284 var InternalError = "InternalError"
 285 var UnauthorizedAccess = "UnauthorizedAccess"
 286 var InvalidRequest = "InvalidRequest"
 287 var SignatureDoesNotMatch = "SignatureDoesNotMatch"
 288
 289 var reRawQueryIndicatesAPI = regexp.MustCompile(`^[a-z]+(&|$)`)
 290
 291 // serveS3 handles r and returns true if r is a request from an S3
 292 // client, otherwise it returns false.
 293 func (h *handler) serveS3(w http.ResponseWriter, r *http.Request) bool {
 294         var token string
 295         if auth := r.Header.Get("Authorization"); strings.HasPrefix(auth, "AWS ") {
 296                 split := strings.SplitN(auth[4:], ":", 2)
 297                 if len(split) < 2 {
 298                         s3ErrorResponse(w, InvalidRequest, "malformed Authorization header", r.URL.Path, http.StatusUnauthorized)
 299                         return true
 300                 }
 301                 token = unescapeKey(split[0])
 302         } else if strings.HasPrefix(auth, s3SignAlgorithm+" ") {
 303                 t, err := h.checks3signature(r)
 304                 if err != nil {
 305                         s3ErrorResponse(w, SignatureDoesNotMatch, "signature verification failed: "+err.Error(), r.URL.Path, http.StatusForbidden)
 306                         return true
 307                 }
 308                 token = t
 309         } else {
 310                 return false
 311         }
 312
 313         fs, sess, tokenUser, err := h.Cache.GetSession(token)
 314         if err != nil {
 315                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 316                 return true
 317         }
 318         readfs := fs
 319         if writeMethod[r.Method] {
 320                 // Create a FileSystem for this request, to avoid
 321                 // exposing incomplete write operations to concurrent
 322                 // requests.
 323                 client := sess.client.WithRequestID(r.Header.Get("X-Request-Id"))
 324                 fs = client.SiteFileSystem(sess.keepclient)
 325                 fs.ForwardSlashNameSubstitution(h.Cluster.Collections.ForwardSlashNameSubstitution)
 326         }
 327
 328         var objectNameGiven bool
 329         var bucketName string
 330         fspath := "/by_id"
 331         if id := arvados.CollectionIDFromDNSName(r.Host); id != "" {
 332                 fspath += "/" + id
 333                 bucketName = id
 334                 objectNameGiven = strings.Count(strings.TrimSuffix(r.URL.Path, "/"), "/") > 0
 335         } else {
 336                 bucketName = strings.SplitN(strings.TrimPrefix(r.URL.Path, "/"), "/", 2)[0]
 337                 objectNameGiven = strings.Count(strings.TrimSuffix(r.URL.Path, "/"), "/") > 1
 338         }
 339         fspath += reMultipleSlashChars.ReplaceAllString(r.URL.Path, "/")
 340
 341         switch {
 342         case r.Method == http.MethodGet && !objectNameGiven:
 343                 // Path is "/{uuid}" or "/{uuid}/", has no object name
 344                 if _, ok := r.URL.Query()["versioning"]; ok {
 345                         // GetBucketVersioning
 346                         w.Header().Set("Content-Type", "application/xml")
 347                         io.WriteString(w, xml.Header)
 348                         fmt.Fprintln(w, `<VersioningConfiguration xmlns="http://s3.amazonaws.com/doc/2006-03-01/"/>`)
 349                 } else if _, ok = r.URL.Query()["location"]; ok {
 350                         // GetBucketLocation
 351                         w.Header().Set("Content-Type", "application/xml")
 352                         io.WriteString(w, xml.Header)
 353                         fmt.Fprintln(w, `<LocationConstraint><LocationConstraint xmlns="http://s3.amazonaws.com/doc/2006-03-01/">`+
 354                                 h.Cluster.ClusterID+
 355                                 `</LocationConstraint></LocationConstraint>`)
 356                 } else if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 357                         // GetBucketWebsite ("GET /bucketid/?website"), GetBucketTagging, etc.
 358                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 359                 } else {
 360                         // ListObjects
 361                         h.s3list(bucketName, w, r, fs)
 362                 }
 363                 return true
 364         case r.Method == http.MethodGet || r.Method == http.MethodHead:
 365                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 366                         // GetObjectRetention ("GET /bucketid/objectid?retention&versionID=..."), etc.
 367                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 368                         return true
 369                 }
 370                 fi, err := fs.Stat(fspath)
 371                 if r.Method == "HEAD" && !objectNameGiven {
 372                         // HeadBucket
 373                         if err == nil && fi.IsDir() {
 374                                 err = setFileInfoHeaders(w.Header(), fs, fspath)
 375                                 if err != nil {
 376                                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 377                                         return true
 378                                 }
 379                                 w.WriteHeader(http.StatusOK)
 380                         } else if os.IsNotExist(err) {
 381                                 s3ErrorResponse(w, NoSuchBucket, "The specified bucket does not exist.", r.URL.Path, http.StatusNotFound)
 382                         } else {
 383                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 384                         }
 385                         return true
 386                 }
 387                 if err == nil && fi.IsDir() && objectNameGiven && strings.HasSuffix(fspath, "/") && h.Cluster.Collections.S3FolderObjects {
 388                         err = setFileInfoHeaders(w.Header(), fs, fspath)
 389                         if err != nil {
 390                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 391                                 return true
 392                         }
 393                         w.Header().Set("Content-Type", "application/x-directory")
 394                         w.WriteHeader(http.StatusOK)
 395                         return true
 396                 }
 397                 if os.IsNotExist(err) ||
 398                         (err != nil && err.Error() == "not a directory") ||
 399                         (fi != nil && fi.IsDir()) {
 400                         s3ErrorResponse(w, NoSuchKey, "The specified key does not exist.", r.URL.Path, http.StatusNotFound)
 401                         return true
 402                 }
 403
 404                 if !h.userPermittedToUploadOrDownload(r.Method, tokenUser) {
 405                         http.Error(w, "Not permitted", http.StatusForbidden)
 406                         return true
 407                 }
 408                 h.logUploadOrDownload(r, sess.arvadosclient, fs, fspath, nil, tokenUser)
 409
 410                 // shallow copy r, and change URL path
 411                 r := *r
 412                 r.URL.Path = fspath
 413                 err = setFileInfoHeaders(w.Header(), fs, fspath)
 414                 if err != nil {
 415                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 416                         return true
 417                 }
 418                 http.FileServer(fs).ServeHTTP(w, &r)
 419                 return true
 420         case r.Method == http.MethodPut:
 421                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 422                         // PutObjectAcl ("PUT /bucketid/objectid?acl&versionID=..."), etc.
 423                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 424                         return true
 425                 }
 426                 if !objectNameGiven {
 427                         s3ErrorResponse(w, InvalidArgument, "Missing object name in PUT request.", r.URL.Path, http.StatusBadRequest)
 428                         return true
 429                 }
 430                 var objectIsDir bool
 431                 if strings.HasSuffix(fspath, "/") {
 432                         if !h.Cluster.Collections.S3FolderObjects {
 433                                 s3ErrorResponse(w, InvalidArgument, "invalid object name: trailing slash", r.URL.Path, http.StatusBadRequest)
 434                                 return true
 435                         }
 436                         n, err := r.Body.Read(make([]byte, 1))
 437                         if err != nil && err != io.EOF {
 438                                 s3ErrorResponse(w, InternalError, fmt.Sprintf("error reading request body: %s", err), r.URL.Path, http.StatusInternalServerError)
 439                                 return true
 440                         } else if n > 0 {
 441                                 s3ErrorResponse(w, InvalidArgument, "cannot create object with trailing '/' char unless content is empty", r.URL.Path, http.StatusBadRequest)
 442                                 return true
 443                         } else if strings.SplitN(r.Header.Get("Content-Type"), ";", 2)[0] != "application/x-directory" {
 444                                 s3ErrorResponse(w, InvalidArgument, "cannot create object with trailing '/' char unless Content-Type is 'application/x-directory'", r.URL.Path, http.StatusBadRequest)
 445                                 return true
 446                         }
 447                         // Given PUT "foo/bar/", we'll use "foo/bar/."
 448                         // in the "ensure parents exist" block below,
 449                         // and then we'll be done.
 450                         fspath += "."
 451                         objectIsDir = true
 452                 }
 453                 fi, err := fs.Stat(fspath)
 454                 if err != nil && err.Error() == "not a directory" {
 455                         // requested foo/bar, but foo is a file
 456                         s3ErrorResponse(w, InvalidArgument, "object name conflicts with existing object", r.URL.Path, http.StatusBadRequest)
 457                         return true
 458                 }
 459                 if strings.HasSuffix(r.URL.Path, "/") && err == nil && !fi.IsDir() {
 460                         // requested foo/bar/, but foo/bar is a file
 461                         s3ErrorResponse(w, InvalidArgument, "object name conflicts with existing object", r.URL.Path, http.StatusBadRequest)
 462                         return true
 463                 }
 464                 // create missing parent/intermediate directories, if any
 465                 for i, c := range fspath {
 466                         if i > 0 && c == '/' {
 467                                 dir := fspath[:i]
 468                                 if strings.HasSuffix(dir, "/") {
 469                                         err = errors.New("invalid object name (consecutive '/' chars)")
 470                                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 471                                         return true
 472                                 }
 473                                 err = fs.Mkdir(dir, 0755)
 474                                 if errors.Is(err, arvados.ErrInvalidArgument) || errors.Is(err, arvados.ErrInvalidOperation) {
 475                                         // Cannot create a directory
 476                                         // here.
 477                                         err = fmt.Errorf("mkdir %q failed: %w", dir, err)
 478                                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 479                                         return true
 480                                 } else if err != nil && !os.IsExist(err) {
 481                                         err = fmt.Errorf("mkdir %q failed: %w", dir, err)
 482                                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 483                                         return true
 484                                 }
 485                         }
 486                 }
 487                 if !objectIsDir {
 488                         f, err := fs.OpenFile(fspath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0644)
 489                         if os.IsNotExist(err) {
 490                                 f, err = fs.OpenFile(fspath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0644)
 491                         }
 492                         if err != nil {
 493                                 err = fmt.Errorf("open %q failed: %w", r.URL.Path, err)
 494                                 s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 495                                 return true
 496                         }
 497                         defer f.Close()
 498
 499                         if !h.userPermittedToUploadOrDownload(r.Method, tokenUser) {
 500                                 http.Error(w, "Not permitted", http.StatusForbidden)
 501                                 return true
 502                         }
 503                         h.logUploadOrDownload(r, sess.arvadosclient, fs, fspath, nil, tokenUser)
 504
 505                         _, err = io.Copy(f, r.Body)
 506                         if err != nil {
 507                                 err = fmt.Errorf("write to %q failed: %w", r.URL.Path, err)
 508                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 509                                 return true
 510                         }
 511                         err = f.Close()
 512                         if err != nil {
 513                                 err = fmt.Errorf("write to %q failed: close: %w", r.URL.Path, err)
 514                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 515                                 return true
 516                         }
 517                 }
 518                 err = h.syncCollection(fs, readfs, fspath)
 519                 if err != nil {
 520                         err = fmt.Errorf("sync failed: %w", err)
 521                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 522                         return true
 523                 }
 524                 w.WriteHeader(http.StatusOK)
 525                 return true
 526         case r.Method == http.MethodDelete:
 527                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 528                         // DeleteObjectTagging ("DELETE /bucketid/objectid?tagging&versionID=..."), etc.
 529                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 530                         return true
 531                 }
 532                 if !objectNameGiven || r.URL.Path == "/" {
 533                         s3ErrorResponse(w, InvalidArgument, "missing object name in DELETE request", r.URL.Path, http.StatusBadRequest)
 534                         return true
 535                 }
 536                 if strings.HasSuffix(fspath, "/") {
 537                         fspath = strings.TrimSuffix(fspath, "/")
 538                         fi, err := fs.Stat(fspath)
 539                         if os.IsNotExist(err) {
 540                                 w.WriteHeader(http.StatusNoContent)
 541                                 return true
 542                         } else if err != nil {
 543                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 544                                 return true
 545                         } else if !fi.IsDir() {
 546                                 // if "foo" exists and is a file, then
 547                                 // "foo/" doesn't exist, so we say
 548                                 // delete was successful.
 549                                 w.WriteHeader(http.StatusNoContent)
 550                                 return true
 551                         }
 552                 } else if fi, err := fs.Stat(fspath); err == nil && fi.IsDir() {
 553                         // if "foo" is a dir, it is visible via S3
 554                         // only as "foo/", not "foo" -- so we leave
 555                         // the dir alone and return 204 to indicate
 556                         // that "foo" does not exist.
 557                         w.WriteHeader(http.StatusNoContent)
 558                         return true
 559                 }
 560                 err = fs.Remove(fspath)
 561                 if os.IsNotExist(err) {
 562                         w.WriteHeader(http.StatusNoContent)
 563                         return true
 564                 }
 565                 if err != nil {
 566                         err = fmt.Errorf("rm failed: %w", err)
 567                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 568                         return true
 569                 }
 570                 err = h.syncCollection(fs, readfs, fspath)
 571                 if err != nil {
 572                         err = fmt.Errorf("sync failed: %w", err)
 573                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 574                         return true
 575                 }
 576                 w.WriteHeader(http.StatusNoContent)
 577                 return true
 578         default:
 579                 s3ErrorResponse(w, InvalidRequest, "method not allowed", r.URL.Path, http.StatusMethodNotAllowed)
 580                 return true
 581         }
 582 }
 583
 584 func (h *handler) syncCollection(srcfs, dstfs arvados.CustomFileSystem, path string) error {
 585         coll, _ := h.determineCollection(srcfs, path)
 586         if coll == nil || coll.UUID == "" {
 587                 return errors.New("could not determine collection to sync")
 588         }
 589         d, err := srcfs.OpenFile("by_id/"+coll.UUID, os.O_RDWR, 0777)
 590         if err != nil {
 591                 return err
 592         }
 593         defer d.Close()
 594         err = d.Sync()
 595         snap, err := d.Snapshot()
 596         if err != nil {
 597                 return err
 598         }
 599         dstd, err := dstfs.OpenFile("by_id/"+coll.UUID, os.O_RDWR, 0777)
 600         if err != nil {
 601                 return err
 602         }
 603         defer dstd.Close()
 604         return dstd.Splice(snap)
 605 }
 606
 607 func setFileInfoHeaders(header http.Header, fs arvados.CustomFileSystem, path string) error {
 608         maybeEncode := func(s string) string {
 609                 for _, c := range s {
 610                         if c > '\u007f' || c < ' ' {
 611                                 return mime.BEncoding.Encode("UTF-8", s)
 612                         }
 613                 }
 614                 return s
 615         }
 616         path = strings.TrimSuffix(path, "/")
 617         var props map[string]interface{}
 618         for {
 619                 fi, err := fs.Stat(path)
 620                 if err != nil {
 621                         return err
 622                 }
 623                 switch src := fi.Sys().(type) {
 624                 case *arvados.Collection:
 625                         props = src.Properties
 626                 case *arvados.Group:
 627                         props = src.Properties
 628                 default:
 629                         if err, ok := src.(error); ok {
 630                                 return err
 631                         }
 632                         // Try parent
 633                         cut := strings.LastIndexByte(path, '/')
 634                         if cut < 0 {
 635                                 return nil
 636                         }
 637                         path = path[:cut]
 638                         continue
 639                 }
 640                 break
 641         }
 642         for k, v := range props {
 643                 if !validMIMEHeaderKey(k) {
 644                         continue
 645                 }
 646                 k = "x-amz-meta-" + k
 647                 if s, ok := v.(string); ok {
 648                         header.Set(k, maybeEncode(s))
 649                 } else if j, err := json.Marshal(v); err == nil {
 650                         header.Set(k, maybeEncode(string(j)))
 651                 }
 652         }
 653         return nil
 654 }
 655
 656 func validMIMEHeaderKey(k string) bool {
 657         check := "z-" + k
 658         return check != textproto.CanonicalMIMEHeaderKey(check)
 659 }
 660
 661 // Call fn on the given path (directory) and its contents, in
 662 // lexicographic order.
 663 //
 664 // If isRoot==true and path is not a directory, return nil.
 665 //
 666 // If fn returns filepath.SkipDir when called on a directory, don't
 667 // descend into that directory.
 668 func walkFS(fs arvados.CustomFileSystem, path string, isRoot bool, fn func(path string, fi os.FileInfo) error) error {
 669         if isRoot {
 670                 fi, err := fs.Stat(path)
 671                 if os.IsNotExist(err) || (err == nil && !fi.IsDir()) {
 672                         return nil
 673                 } else if err != nil {
 674                         return err
 675                 }
 676                 err = fn(path, fi)
 677                 if err == filepath.SkipDir {
 678                         return nil
 679                 } else if err != nil {
 680                         return err
 681                 }
 682         }
 683         f, err := fs.Open(path)
 684         if os.IsNotExist(err) && isRoot {
 685                 return nil
 686         } else if err != nil {
 687                 return fmt.Errorf("open %q: %w", path, err)
 688         }
 689         defer f.Close()
 690         if path == "/" {
 691                 path = ""
 692         }
 693         fis, err := f.Readdir(-1)
 694         if err != nil {
 695                 return err
 696         }
 697         sort.Slice(fis, func(i, j int) bool { return fis[i].Name() < fis[j].Name() })
 698         for _, fi := range fis {
 699                 err = fn(path+"/"+fi.Name(), fi)
 700                 if err == filepath.SkipDir {
 701                         continue
 702                 } else if err != nil {
 703                         return err
 704                 }
 705                 if fi.IsDir() {
 706                         err = walkFS(fs, path+"/"+fi.Name(), false, fn)
 707                         if err != nil {
 708                                 return err
 709                         }
 710                 }
 711         }
 712         return nil
 713 }
 714
 715 var errDone = errors.New("done")
 716
 717 func (h *handler) s3list(bucket string, w http.ResponseWriter, r *http.Request, fs arvados.CustomFileSystem) {
 718         var params struct {
 719                 v2                bool
 720                 delimiter         string
 721                 maxKeys           int
 722                 prefix            string
 723                 marker            string // decoded continuationToken (v2) or provided by client (v1)
 724                 startAfter        string // v2
 725                 continuationToken string // v2
 726                 encodingTypeURL   bool   // v2
 727         }
 728         params.delimiter = r.FormValue("delimiter")
 729         if mk, _ := strconv.ParseInt(r.FormValue("max-keys"), 10, 64); mk > 0 && mk < s3MaxKeys {
 730                 params.maxKeys = int(mk)
 731         } else {
 732                 params.maxKeys = s3MaxKeys
 733         }
 734         params.prefix = r.FormValue("prefix")
 735         switch r.FormValue("list-type") {
 736         case "":
 737         case "2":
 738                 params.v2 = true
 739         default:
 740                 http.Error(w, "invalid list-type parameter", http.StatusBadRequest)
 741                 return
 742         }
 743         if params.v2 {
 744                 params.continuationToken = r.FormValue("continuation-token")
 745                 marker, err := base64.StdEncoding.DecodeString(params.continuationToken)
 746                 if err != nil {
 747                         http.Error(w, "invalid continuation token", http.StatusBadRequest)
 748                         return
 749                 }
 750                 params.marker = string(marker)
 751                 params.startAfter = r.FormValue("start-after")
 752                 switch r.FormValue("encoding-type") {
 753                 case "":
 754                 case "url":
 755                         params.encodingTypeURL = true
 756                 default:
 757                         http.Error(w, "invalid encoding-type parameter", http.StatusBadRequest)
 758                         return
 759                 }
 760         } else {
 761                 params.marker = r.FormValue("marker")
 762         }
 763
 764         bucketdir := "by_id/" + bucket
 765         // walkpath is the directory (relative to bucketdir) we need
 766         // to walk: the innermost directory that is guaranteed to
 767         // contain all paths that have the requested prefix. Examples:
 768         // prefix "foo/bar"  => walkpath "foo"
 769         // prefix "foo/bar/" => walkpath "foo/bar"
 770         // prefix "foo"      => walkpath ""
 771         // prefix ""         => walkpath ""
 772         walkpath := params.prefix
 773         if cut := strings.LastIndex(walkpath, "/"); cut >= 0 {
 774                 walkpath = walkpath[:cut]
 775         } else {
 776                 walkpath = ""
 777         }
 778
 779         resp := listV2Resp{
 780                 Name:              bucket,
 781                 Prefix:            params.prefix,
 782                 Delimiter:         params.delimiter,
 783                 MaxKeys:           params.maxKeys,
 784                 ContinuationToken: r.FormValue("continuation-token"),
 785                 StartAfter:        params.startAfter,
 786         }
 787         nextMarker := ""
 788
 789         commonPrefixes := map[string]bool{}
 790         err := walkFS(fs, strings.TrimSuffix(bucketdir+"/"+walkpath, "/"), true, func(path string, fi os.FileInfo) error {
 791                 if path == bucketdir {
 792                         return nil
 793                 }
 794                 path = path[len(bucketdir)+1:]
 795                 filesize := fi.Size()
 796                 if fi.IsDir() {
 797                         path += "/"
 798                         filesize = 0
 799                 }
 800                 if len(path) <= len(params.prefix) {
 801                         if path > params.prefix[:len(path)] {
 802                                 // with prefix "foobar", walking "fooz" means we're done
 803                                 return errDone
 804                         }
 805                         if path < params.prefix[:len(path)] {
 806                                 // with prefix "foobar", walking "foobag" is pointless
 807                                 return filepath.SkipDir
 808                         }
 809                         if fi.IsDir() && !strings.HasPrefix(params.prefix+"/", path) {
 810                                 // with prefix "foo/bar", walking "fo"
 811                                 // is pointless (but walking "foo" or
 812                                 // "foo/bar" is necessary)
 813                                 return filepath.SkipDir
 814                         }
 815                         if len(path) < len(params.prefix) {
 816                                 // can't skip anything, and this entry
 817                                 // isn't in the results, so just
 818                                 // continue descent
 819                                 return nil
 820                         }
 821                 } else {
 822                         if path[:len(params.prefix)] > params.prefix {
 823                                 // with prefix "foobar", nothing we
 824                                 // see after "foozzz" is relevant
 825                                 return errDone
 826                         }
 827                 }
 828                 if path < params.marker || path < params.prefix || path <= params.startAfter {
 829                         return nil
 830                 }
 831                 if fi.IsDir() && !h.Cluster.Collections.S3FolderObjects {
 832                         // Note we don't add anything to
 833                         // commonPrefixes here even if delimiter is
 834                         // "/". We descend into the directory, and
 835                         // return a commonPrefix only if we end up
 836                         // finding a regular file inside it.
 837                         return nil
 838                 }
 839                 if len(resp.Contents)+len(commonPrefixes) >= params.maxKeys {
 840                         resp.IsTruncated = true
 841                         if params.delimiter != "" || params.v2 {
 842                                 nextMarker = path
 843                         }
 844                         return errDone
 845                 }
 846                 if params.delimiter != "" {
 847                         idx := strings.Index(path[len(params.prefix):], params.delimiter)
 848                         if idx >= 0 {
 849                                 // with prefix "foobar" and delimiter
 850                                 // "z", when we hit "foobar/baz", we
 851                                 // add "/baz" to commonPrefixes and
 852                                 // stop descending.
 853                                 commonPrefixes[path[:len(params.prefix)+idx+1]] = true
 854                                 return filepath.SkipDir
 855                         }
 856                 }
 857                 resp.Contents = append(resp.Contents, s3.Key{
 858                         Key:          path,
 859                         LastModified: fi.ModTime().UTC().Format("2006-01-02T15:04:05.999") + "Z",
 860                         Size:         filesize,
 861                 })
 862                 return nil
 863         })
 864         if err != nil && err != errDone {
 865                 http.Error(w, err.Error(), http.StatusInternalServerError)
 866                 return
 867         }
 868         if params.delimiter != "" {
 869                 resp.CommonPrefixes = make([]commonPrefix, 0, len(commonPrefixes))
 870                 for prefix := range commonPrefixes {
 871                         resp.CommonPrefixes = append(resp.CommonPrefixes, commonPrefix{prefix})
 872                 }
 873                 sort.Slice(resp.CommonPrefixes, func(i, j int) bool { return resp.CommonPrefixes[i].Prefix < resp.CommonPrefixes[j].Prefix })
 874         }
 875         resp.KeyCount = len(resp.Contents)
 876         var respV1orV2 interface{}
 877
 878         if params.encodingTypeURL {
 879                 // https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html
 880                 // "If you specify the encoding-type request
 881                 // parameter, Amazon S3 includes this element in the
 882                 // response, and returns encoded key name values in
 883                 // the following response elements:
 884                 //
 885                 // Delimiter, Prefix, Key, and StartAfter.
 886                 //
 887                 //      Type: String
 888                 //
 889                 // Valid Values: url"
 890                 //
 891                 // This is somewhat vague but in practice it appears
 892                 // to mean x-www-form-urlencoded as in RFC1866 8.2.1
 893                 // para 1 (encode space as "+") rather than straight
 894                 // percent-encoding as in RFC1738 2.2.  Presumably,
 895                 // the intent is to allow the client to decode XML and
 896                 // then paste the strings directly into another URI
 897                 // query or POST form like "https://host/path?foo=" +
 898                 // foo + "&bar=" + bar.
 899                 resp.EncodingType = "url"
 900                 resp.Delimiter = url.QueryEscape(resp.Delimiter)
 901                 resp.Prefix = url.QueryEscape(resp.Prefix)
 902                 resp.StartAfter = url.QueryEscape(resp.StartAfter)
 903                 for i, ent := range resp.Contents {
 904                         ent.Key = url.QueryEscape(ent.Key)
 905                         resp.Contents[i] = ent
 906                 }
 907                 for i, ent := range resp.CommonPrefixes {
 908                         ent.Prefix = url.QueryEscape(ent.Prefix)
 909                         resp.CommonPrefixes[i] = ent
 910                 }
 911         }
 912
 913         if params.v2 {
 914                 resp.NextContinuationToken = base64.StdEncoding.EncodeToString([]byte(nextMarker))
 915                 respV1orV2 = resp
 916         } else {
 917                 respV1orV2 = listV1Resp{
 918                         CommonPrefixes: resp.CommonPrefixes,
 919                         NextMarker:     nextMarker,
 920                         KeyCount:       resp.KeyCount,
 921                         ListResp: s3.ListResp{
 922                                 IsTruncated: resp.IsTruncated,
 923                                 Name:        bucket,
 924                                 Prefix:      params.prefix,
 925                                 Delimiter:   params.delimiter,
 926                                 Marker:      params.marker,
 927                                 MaxKeys:     params.maxKeys,
 928                                 Contents:    resp.Contents,
 929                         },
 930                 }
 931         }
 932
 933         w.Header().Set("Content-Type", "application/xml")
 934         io.WriteString(w, xml.Header)
 935         if err := xml.NewEncoder(w).Encode(respV1orV2); err != nil {
 936                 ctxlog.FromContext(r.Context()).WithError(err).Error("error writing xml response")
 937         }
 938 }