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