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         normalizedURL := *r.URL
 140         normalizedURL.RawPath = ""
 141         normalizedURL.Path = reMultipleSlashChars.ReplaceAllString(normalizedURL.Path, "/")
 142         ctxlog.FromContext(r.Context()).Infof("escapedPath %s", normalizedURL.EscapedPath())
 143         canonicalRequest := fmt.Sprintf("%s\n%s\n%s\n%s\n%s\n%s", r.Method, normalizedURL.EscapedPath(), s3querystring(r.URL), canonicalHeaders, signedHeaders, r.Header.Get("X-Amz-Content-Sha256"))
 144         ctxlog.FromContext(r.Context()).Debugf("s3stringToSign: canonicalRequest %s", canonicalRequest)
 145         return fmt.Sprintf("%s\n%s\n%s\n%s", alg, r.Header.Get("X-Amz-Date"), scope, hashdigest(sha256.New(), canonicalRequest)), nil
 146 }
 147
 148 func s3signature(secretKey, scope, signedHeaders, stringToSign string) (string, error) {
 149         // scope is {datestamp}/{region}/{service}/aws4_request
 150         drs := strings.Split(scope, "/")
 151         if len(drs) != 4 {
 152                 return "", fmt.Errorf("invalid scope %q", scope)
 153         }
 154         key := s3signatureKey(secretKey, drs[0], drs[1], drs[2])
 155         return hashdigest(hmac.New(sha256.New, key), stringToSign), nil
 156 }
 157
 158 var v2tokenUnderscore = regexp.MustCompile(`^v2_[a-z0-9]{5}-gj3su-[a-z0-9]{15}_`)
 159
 160 func unescapeKey(key string) string {
 161         if v2tokenUnderscore.MatchString(key) {
 162                 // Entire Arvados token, with "/" replaced by "_" to
 163                 // avoid colliding with the Authorization header
 164                 // format.
 165                 return strings.Replace(key, "_", "/", -1)
 166         } else if s, err := url.PathUnescape(key); err == nil {
 167                 return s
 168         } else {
 169                 return key
 170         }
 171 }
 172
 173 // checks3signature verifies the given S3 V4 signature and returns the
 174 // Arvados token that corresponds to the given accessKey. An error is
 175 // returned if accessKey is not a valid token UUID or the signature
 176 // does not match.
 177 func (h *handler) checks3signature(r *http.Request) (string, error) {
 178         var key, scope, signedHeaders, signature string
 179         authstring := strings.TrimPrefix(r.Header.Get("Authorization"), s3SignAlgorithm+" ")
 180         for _, cmpt := range strings.Split(authstring, ",") {
 181                 cmpt = strings.TrimSpace(cmpt)
 182                 split := strings.SplitN(cmpt, "=", 2)
 183                 switch {
 184                 case len(split) != 2:
 185                         // (?) ignore
 186                 case split[0] == "Credential":
 187                         keyandscope := strings.SplitN(split[1], "/", 2)
 188                         if len(keyandscope) == 2 {
 189                                 key, scope = keyandscope[0], keyandscope[1]
 190                         }
 191                 case split[0] == "SignedHeaders":
 192                         signedHeaders = split[1]
 193                 case split[0] == "Signature":
 194                         signature = split[1]
 195                 }
 196         }
 197
 198         client := (&arvados.Client{
 199                 APIHost:  h.Config.cluster.Services.Controller.ExternalURL.Host,
 200                 Insecure: h.Config.cluster.TLS.Insecure,
 201         }).WithRequestID(r.Header.Get("X-Request-Id"))
 202         var aca arvados.APIClientAuthorization
 203         var secret string
 204         var err error
 205         if len(key) == 27 && key[5:12] == "-gj3su-" {
 206                 // Access key is the UUID of an Arvados token, secret
 207                 // key is the secret part.
 208                 ctx := arvados.ContextWithAuthorization(r.Context(), "Bearer "+h.Config.cluster.SystemRootToken)
 209                 err = client.RequestAndDecodeContext(ctx, &aca, "GET", "arvados/v1/api_client_authorizations/"+key, nil, nil)
 210                 secret = aca.APIToken
 211         } else {
 212                 // Access key and secret key are both an entire
 213                 // Arvados token or OIDC access token.
 214                 ctx := arvados.ContextWithAuthorization(r.Context(), "Bearer "+unescapeKey(key))
 215                 err = client.RequestAndDecodeContext(ctx, &aca, "GET", "arvados/v1/api_client_authorizations/current", nil, nil)
 216                 secret = key
 217         }
 218         if err != nil {
 219                 ctxlog.FromContext(r.Context()).WithError(err).WithField("UUID", key).Info("token lookup failed")
 220                 return "", errors.New("invalid access key")
 221         }
 222         stringToSign, err := s3stringToSign(s3SignAlgorithm, scope, signedHeaders, r)
 223         if err != nil {
 224                 return "", err
 225         }
 226         expect, err := s3signature(secret, scope, signedHeaders, stringToSign)
 227         if err != nil {
 228                 return "", err
 229         } else if expect != signature {
 230                 return "", fmt.Errorf("signature does not match (scope %q signedHeaders %q stringToSign %q)", scope, signedHeaders, stringToSign)
 231         }
 232         return aca.TokenV2(), nil
 233 }
 234
 235 func s3ErrorResponse(w http.ResponseWriter, s3code string, message string, resource string, code int) {
 236         w.Header().Set("Content-Type", "application/xml")
 237         w.Header().Set("X-Content-Type-Options", "nosniff")
 238         w.WriteHeader(code)
 239         var errstruct struct {
 240                 Code      string
 241                 Message   string
 242                 Resource  string
 243                 RequestId string
 244         }
 245         errstruct.Code = s3code
 246         errstruct.Message = message
 247         errstruct.Resource = resource
 248         errstruct.RequestId = ""
 249         enc := xml.NewEncoder(w)
 250         fmt.Fprint(w, xml.Header)
 251         enc.EncodeElement(errstruct, xml.StartElement{Name: xml.Name{Local: "Error"}})
 252 }
 253
 254 var NoSuchKey = "NoSuchKey"
 255 var NoSuchBucket = "NoSuchBucket"
 256 var InvalidArgument = "InvalidArgument"
 257 var InternalError = "InternalError"
 258 var UnauthorizedAccess = "UnauthorizedAccess"
 259 var InvalidRequest = "InvalidRequest"
 260 var SignatureDoesNotMatch = "SignatureDoesNotMatch"
 261
 262 var reRawQueryIndicatesAPI = regexp.MustCompile(`^[a-z]+(&|$)`)
 263
 264 // serveS3 handles r and returns true if r is a request from an S3
 265 // client, otherwise it returns false.
 266 func (h *handler) serveS3(w http.ResponseWriter, r *http.Request) bool {
 267         var token string
 268         if auth := r.Header.Get("Authorization"); strings.HasPrefix(auth, "AWS ") {
 269                 split := strings.SplitN(auth[4:], ":", 2)
 270                 if len(split) < 2 {
 271                         s3ErrorResponse(w, InvalidRequest, "malformed Authorization header", r.URL.Path, http.StatusUnauthorized)
 272                         return true
 273                 }
 274                 token = unescapeKey(split[0])
 275         } else if strings.HasPrefix(auth, s3SignAlgorithm+" ") {
 276                 t, err := h.checks3signature(r)
 277                 if err != nil {
 278                         s3ErrorResponse(w, SignatureDoesNotMatch, "signature verification failed: "+err.Error(), r.URL.Path, http.StatusForbidden)
 279                         return true
 280                 }
 281                 token = t
 282         } else {
 283                 return false
 284         }
 285
 286         var err error
 287         var fs arvados.CustomFileSystem
 288         if r.Method == http.MethodGet || r.Method == http.MethodHead {
 289                 // Use a single session (cached FileSystem) across
 290                 // multiple read requests.
 291                 fs, err = h.Config.Cache.GetSession(token)
 292                 if err != nil {
 293                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 294                         return true
 295                 }
 296         } else {
 297                 // Create a FileSystem for this request, to avoid
 298                 // exposing incomplete write operations to concurrent
 299                 // requests.
 300                 _, kc, client, release, err := h.getClients(r.Header.Get("X-Request-Id"), token)
 301                 if err != nil {
 302                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 303                         return true
 304                 }
 305                 defer release()
 306                 fs = client.SiteFileSystem(kc)
 307                 fs.ForwardSlashNameSubstitution(h.Config.cluster.Collections.ForwardSlashNameSubstitution)
 308         }
 309
 310         var objectNameGiven bool
 311         var bucketName string
 312         fspath := "/by_id"
 313         if id := parseCollectionIDFromDNSName(r.Host); id != "" {
 314                 fspath += "/" + id
 315                 bucketName = id
 316                 objectNameGiven = strings.Count(strings.TrimSuffix(r.URL.Path, "/"), "/") > 0
 317         } else {
 318                 bucketName = strings.SplitN(strings.TrimPrefix(r.URL.Path, "/"), "/", 2)[0]
 319                 objectNameGiven = strings.Count(strings.TrimSuffix(r.URL.Path, "/"), "/") > 1
 320         }
 321         fspath += reMultipleSlashChars.ReplaceAllString(r.URL.Path, "/")
 322
 323         switch {
 324         case r.Method == http.MethodGet && !objectNameGiven:
 325                 // Path is "/{uuid}" or "/{uuid}/", has no object name
 326                 if _, ok := r.URL.Query()["versioning"]; ok {
 327                         // GetBucketVersioning
 328                         w.Header().Set("Content-Type", "application/xml")
 329                         io.WriteString(w, xml.Header)
 330                         fmt.Fprintln(w, `<VersioningConfiguration xmlns="http://s3.amazonaws.com/doc/2006-03-01/"/>`)
 331                 } else if _, ok = r.URL.Query()["location"]; ok {
 332                         // GetBucketLocation
 333                         w.Header().Set("Content-Type", "application/xml")
 334                         io.WriteString(w, xml.Header)
 335                         fmt.Fprintln(w, `<LocationConstraint><LocationConstraint xmlns="http://s3.amazonaws.com/doc/2006-03-01/">`+
 336                                 h.Config.cluster.ClusterID+
 337                                 `</LocationConstraint></LocationConstraint>`)
 338                 } else if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 339                         // GetBucketWebsite ("GET /bucketid/?website"), GetBucketTagging, etc.
 340                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 341                 } else {
 342                         // ListObjects
 343                         h.s3list(bucketName, w, r, fs)
 344                 }
 345                 return true
 346         case r.Method == http.MethodGet || r.Method == http.MethodHead:
 347                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 348                         // GetObjectRetention ("GET /bucketid/objectid?retention&versionID=..."), etc.
 349                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 350                         return true
 351                 }
 352                 fi, err := fs.Stat(fspath)
 353                 if r.Method == "HEAD" && !objectNameGiven {
 354                         // HeadBucket
 355                         if err == nil && fi.IsDir() {
 356                                 w.WriteHeader(http.StatusOK)
 357                         } else if os.IsNotExist(err) {
 358                                 s3ErrorResponse(w, NoSuchBucket, "The specified bucket does not exist.", r.URL.Path, http.StatusNotFound)
 359                         } else {
 360                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 361                         }
 362                         return true
 363                 }
 364                 if err == nil && fi.IsDir() && objectNameGiven && strings.HasSuffix(fspath, "/") && h.Config.cluster.Collections.S3FolderObjects {
 365                         w.Header().Set("Content-Type", "application/x-directory")
 366                         w.WriteHeader(http.StatusOK)
 367                         return true
 368                 }
 369                 if os.IsNotExist(err) ||
 370                         (err != nil && err.Error() == "not a directory") ||
 371                         (fi != nil && fi.IsDir()) {
 372                         s3ErrorResponse(w, NoSuchKey, "The specified key does not exist.", r.URL.Path, http.StatusNotFound)
 373                         return true
 374                 }
 375                 // shallow copy r, and change URL path
 376                 r := *r
 377                 r.URL.Path = fspath
 378                 http.FileServer(fs).ServeHTTP(w, &r)
 379                 return true
 380         case r.Method == http.MethodPut:
 381                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 382                         // PutObjectAcl ("PUT /bucketid/objectid?acl&versionID=..."), etc.
 383                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 384                         return true
 385                 }
 386                 if !objectNameGiven {
 387                         s3ErrorResponse(w, InvalidArgument, "Missing object name in PUT request.", r.URL.Path, http.StatusBadRequest)
 388                         return true
 389                 }
 390                 var objectIsDir bool
 391                 if strings.HasSuffix(fspath, "/") {
 392                         if !h.Config.cluster.Collections.S3FolderObjects {
 393                                 s3ErrorResponse(w, InvalidArgument, "invalid object name: trailing slash", r.URL.Path, http.StatusBadRequest)
 394                                 return true
 395                         }
 396                         n, err := r.Body.Read(make([]byte, 1))
 397                         if err != nil && err != io.EOF {
 398                                 s3ErrorResponse(w, InternalError, fmt.Sprintf("error reading request body: %s", err), r.URL.Path, http.StatusInternalServerError)
 399                                 return true
 400                         } else if n > 0 {
 401                                 s3ErrorResponse(w, InvalidArgument, "cannot create object with trailing '/' char unless content is empty", r.URL.Path, http.StatusBadRequest)
 402                                 return true
 403                         } else if strings.SplitN(r.Header.Get("Content-Type"), ";", 2)[0] != "application/x-directory" {
 404                                 s3ErrorResponse(w, InvalidArgument, "cannot create object with trailing '/' char unless Content-Type is 'application/x-directory'", r.URL.Path, http.StatusBadRequest)
 405                                 return true
 406                         }
 407                         // Given PUT "foo/bar/", we'll use "foo/bar/."
 408                         // in the "ensure parents exist" block below,
 409                         // and then we'll be done.
 410                         fspath += "."
 411                         objectIsDir = true
 412                 }
 413                 fi, err := fs.Stat(fspath)
 414                 if err != nil && err.Error() == "not a directory" {
 415                         // requested foo/bar, but foo is a file
 416                         s3ErrorResponse(w, InvalidArgument, "object name conflicts with existing object", r.URL.Path, http.StatusBadRequest)
 417                         return true
 418                 }
 419                 if strings.HasSuffix(r.URL.Path, "/") && err == nil && !fi.IsDir() {
 420                         // requested foo/bar/, but foo/bar is a file
 421                         s3ErrorResponse(w, InvalidArgument, "object name conflicts with existing object", r.URL.Path, http.StatusBadRequest)
 422                         return true
 423                 }
 424                 // create missing parent/intermediate directories, if any
 425                 for i, c := range fspath {
 426                         if i > 0 && c == '/' {
 427                                 dir := fspath[:i]
 428                                 if strings.HasSuffix(dir, "/") {
 429                                         err = errors.New("invalid object name (consecutive '/' chars)")
 430                                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 431                                         return true
 432                                 }
 433                                 err = fs.Mkdir(dir, 0755)
 434                                 if err == arvados.ErrInvalidArgument {
 435                                         // Cannot create a directory
 436                                         // here.
 437                                         err = fmt.Errorf("mkdir %q failed: %w", dir, err)
 438                                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 439                                         return true
 440                                 } else if err != nil && !os.IsExist(err) {
 441                                         err = fmt.Errorf("mkdir %q failed: %w", dir, err)
 442                                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 443                                         return true
 444                                 }
 445                         }
 446                 }
 447                 if !objectIsDir {
 448                         f, err := fs.OpenFile(fspath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0644)
 449                         if os.IsNotExist(err) {
 450                                 f, err = fs.OpenFile(fspath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0644)
 451                         }
 452                         if err != nil {
 453                                 err = fmt.Errorf("open %q failed: %w", r.URL.Path, err)
 454                                 s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 455                                 return true
 456                         }
 457                         defer f.Close()
 458                         _, err = io.Copy(f, r.Body)
 459                         if err != nil {
 460                                 err = fmt.Errorf("write to %q failed: %w", r.URL.Path, err)
 461                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 462                                 return true
 463                         }
 464                         err = f.Close()
 465                         if err != nil {
 466                                 err = fmt.Errorf("write to %q failed: close: %w", r.URL.Path, err)
 467                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 468                                 return true
 469                         }
 470                 }
 471                 err = fs.Sync()
 472                 if err != nil {
 473                         err = fmt.Errorf("sync failed: %w", err)
 474                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 475                         return true
 476                 }
 477                 // Ensure a subsequent read operation will see the changes.
 478                 h.Config.Cache.ResetSession(token)
 479                 w.WriteHeader(http.StatusOK)
 480                 return true
 481         case r.Method == http.MethodDelete:
 482                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 483                         // DeleteObjectTagging ("DELETE /bucketid/objectid?tagging&versionID=..."), etc.
 484                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 485                         return true
 486                 }
 487                 if !objectNameGiven || r.URL.Path == "/" {
 488                         s3ErrorResponse(w, InvalidArgument, "missing object name in DELETE request", r.URL.Path, http.StatusBadRequest)
 489                         return true
 490                 }
 491                 if strings.HasSuffix(fspath, "/") {
 492                         fspath = strings.TrimSuffix(fspath, "/")
 493                         fi, err := fs.Stat(fspath)
 494                         if os.IsNotExist(err) {
 495                                 w.WriteHeader(http.StatusNoContent)
 496                                 return true
 497                         } else if err != nil {
 498                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 499                                 return true
 500                         } else if !fi.IsDir() {
 501                                 // if "foo" exists and is a file, then
 502                                 // "foo/" doesn't exist, so we say
 503                                 // delete was successful.
 504                                 w.WriteHeader(http.StatusNoContent)
 505                                 return true
 506                         }
 507                 } else if fi, err := fs.Stat(fspath); err == nil && fi.IsDir() {
 508                         // if "foo" is a dir, it is visible via S3
 509                         // only as "foo/", not "foo" -- so we leave
 510                         // the dir alone and return 204 to indicate
 511                         // that "foo" does not exist.
 512                         w.WriteHeader(http.StatusNoContent)
 513                         return true
 514                 }
 515                 err = fs.Remove(fspath)
 516                 if os.IsNotExist(err) {
 517                         w.WriteHeader(http.StatusNoContent)
 518                         return true
 519                 }
 520                 if err != nil {
 521                         err = fmt.Errorf("rm failed: %w", err)
 522                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 523                         return true
 524                 }
 525                 err = fs.Sync()
 526                 if err != nil {
 527                         err = fmt.Errorf("sync failed: %w", err)
 528                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 529                         return true
 530                 }
 531                 // Ensure a subsequent read operation will see the changes.
 532                 h.Config.Cache.ResetSession(token)
 533                 w.WriteHeader(http.StatusNoContent)
 534                 return true
 535         default:
 536                 s3ErrorResponse(w, InvalidRequest, "method not allowed", r.URL.Path, http.StatusMethodNotAllowed)
 537                 return true
 538         }
 539 }
 540
 541 // Call fn on the given path (directory) and its contents, in
 542 // lexicographic order.
 543 //
 544 // If isRoot==true and path is not a directory, return nil.
 545 //
 546 // If fn returns filepath.SkipDir when called on a directory, don't
 547 // descend into that directory.
 548 func walkFS(fs arvados.CustomFileSystem, path string, isRoot bool, fn func(path string, fi os.FileInfo) error) error {
 549         if isRoot {
 550                 fi, err := fs.Stat(path)
 551                 if os.IsNotExist(err) || (err == nil && !fi.IsDir()) {
 552                         return nil
 553                 } else if err != nil {
 554                         return err
 555                 }
 556                 err = fn(path, fi)
 557                 if err == filepath.SkipDir {
 558                         return nil
 559                 } else if err != nil {
 560                         return err
 561                 }
 562         }
 563         f, err := fs.Open(path)
 564         if os.IsNotExist(err) && isRoot {
 565                 return nil
 566         } else if err != nil {
 567                 return fmt.Errorf("open %q: %w", path, err)
 568         }
 569         defer f.Close()
 570         if path == "/" {
 571                 path = ""
 572         }
 573         fis, err := f.Readdir(-1)
 574         if err != nil {
 575                 return err
 576         }
 577         sort.Slice(fis, func(i, j int) bool { return fis[i].Name() < fis[j].Name() })
 578         for _, fi := range fis {
 579                 err = fn(path+"/"+fi.Name(), fi)
 580                 if err == filepath.SkipDir {
 581                         continue
 582                 } else if err != nil {
 583                         return err
 584                 }
 585                 if fi.IsDir() {
 586                         err = walkFS(fs, path+"/"+fi.Name(), false, fn)
 587                         if err != nil {
 588                                 return err
 589                         }
 590                 }
 591         }
 592         return nil
 593 }
 594
 595 var errDone = errors.New("done")
 596
 597 func (h *handler) s3list(bucket string, w http.ResponseWriter, r *http.Request, fs arvados.CustomFileSystem) {
 598         var params struct {
 599                 v2                bool
 600                 delimiter         string
 601                 maxKeys           int
 602                 prefix            string
 603                 marker            string // decoded continuationToken (v2) or provided by client (v1)
 604                 startAfter        string // v2
 605                 continuationToken string // v2
 606                 encodingTypeURL   bool   // v2
 607         }
 608         params.delimiter = r.FormValue("delimiter")
 609         if mk, _ := strconv.ParseInt(r.FormValue("max-keys"), 10, 64); mk > 0 && mk < s3MaxKeys {
 610                 params.maxKeys = int(mk)
 611         } else {
 612                 params.maxKeys = s3MaxKeys
 613         }
 614         params.prefix = r.FormValue("prefix")
 615         switch r.FormValue("list-type") {
 616         case "":
 617         case "2":
 618                 params.v2 = true
 619         default:
 620                 http.Error(w, "invalid list-type parameter", http.StatusBadRequest)
 621                 return
 622         }
 623         if params.v2 {
 624                 params.continuationToken = r.FormValue("continuation-token")
 625                 marker, err := base64.StdEncoding.DecodeString(params.continuationToken)
 626                 if err != nil {
 627                         http.Error(w, "invalid continuation token", http.StatusBadRequest)
 628                         return
 629                 }
 630                 params.marker = string(marker)
 631                 params.startAfter = r.FormValue("start-after")
 632                 switch r.FormValue("encoding-type") {
 633                 case "":
 634                 case "url":
 635                         params.encodingTypeURL = true
 636                 default:
 637                         http.Error(w, "invalid encoding-type parameter", http.StatusBadRequest)
 638                         return
 639                 }
 640         } else {
 641                 params.marker = r.FormValue("marker")
 642         }
 643
 644         bucketdir := "by_id/" + bucket
 645         // walkpath is the directory (relative to bucketdir) we need
 646         // to walk: the innermost directory that is guaranteed to
 647         // contain all paths that have the requested prefix. Examples:
 648         // prefix "foo/bar"  => walkpath "foo"
 649         // prefix "foo/bar/" => walkpath "foo/bar"
 650         // prefix "foo"      => walkpath ""
 651         // prefix ""         => walkpath ""
 652         walkpath := params.prefix
 653         if cut := strings.LastIndex(walkpath, "/"); cut >= 0 {
 654                 walkpath = walkpath[:cut]
 655         } else {
 656                 walkpath = ""
 657         }
 658
 659         resp := listV2Resp{
 660                 Name:              bucket,
 661                 Prefix:            params.prefix,
 662                 Delimiter:         params.delimiter,
 663                 MaxKeys:           params.maxKeys,
 664                 ContinuationToken: r.FormValue("continuation-token"),
 665                 StartAfter:        params.startAfter,
 666         }
 667         nextMarker := ""
 668
 669         commonPrefixes := map[string]bool{}
 670         err := walkFS(fs, strings.TrimSuffix(bucketdir+"/"+walkpath, "/"), true, func(path string, fi os.FileInfo) error {
 671                 if path == bucketdir {
 672                         return nil
 673                 }
 674                 path = path[len(bucketdir)+1:]
 675                 filesize := fi.Size()
 676                 if fi.IsDir() {
 677                         path += "/"
 678                         filesize = 0
 679                 }
 680                 if len(path) <= len(params.prefix) {
 681                         if path > params.prefix[:len(path)] {
 682                                 // with prefix "foobar", walking "fooz" means we're done
 683                                 return errDone
 684                         }
 685                         if path < params.prefix[:len(path)] {
 686                                 // with prefix "foobar", walking "foobag" is pointless
 687                                 return filepath.SkipDir
 688                         }
 689                         if fi.IsDir() && !strings.HasPrefix(params.prefix+"/", path) {
 690                                 // with prefix "foo/bar", walking "fo"
 691                                 // is pointless (but walking "foo" or
 692                                 // "foo/bar" is necessary)
 693                                 return filepath.SkipDir
 694                         }
 695                         if len(path) < len(params.prefix) {
 696                                 // can't skip anything, and this entry
 697                                 // isn't in the results, so just
 698                                 // continue descent
 699                                 return nil
 700                         }
 701                 } else {
 702                         if path[:len(params.prefix)] > params.prefix {
 703                                 // with prefix "foobar", nothing we
 704                                 // see after "foozzz" is relevant
 705                                 return errDone
 706                         }
 707                 }
 708                 if path < params.marker || path < params.prefix || path <= params.startAfter {
 709                         return nil
 710                 }
 711                 if fi.IsDir() && !h.Config.cluster.Collections.S3FolderObjects {
 712                         // Note we don't add anything to
 713                         // commonPrefixes here even if delimiter is
 714                         // "/". We descend into the directory, and
 715                         // return a commonPrefix only if we end up
 716                         // finding a regular file inside it.
 717                         return nil
 718                 }
 719                 if len(resp.Contents)+len(commonPrefixes) >= params.maxKeys {
 720                         resp.IsTruncated = true
 721                         if params.delimiter != "" || params.v2 {
 722                                 nextMarker = path
 723                         }
 724                         return errDone
 725                 }
 726                 if params.delimiter != "" {
 727                         idx := strings.Index(path[len(params.prefix):], params.delimiter)
 728                         if idx >= 0 {
 729                                 // with prefix "foobar" and delimiter
 730                                 // "z", when we hit "foobar/baz", we
 731                                 // add "/baz" to commonPrefixes and
 732                                 // stop descending.
 733                                 commonPrefixes[path[:len(params.prefix)+idx+1]] = true
 734                                 return filepath.SkipDir
 735                         }
 736                 }
 737                 resp.Contents = append(resp.Contents, s3.Key{
 738                         Key:          path,
 739                         LastModified: fi.ModTime().UTC().Format("2006-01-02T15:04:05.999") + "Z",
 740                         Size:         filesize,
 741                 })
 742                 return nil
 743         })
 744         if err != nil && err != errDone {
 745                 http.Error(w, err.Error(), http.StatusInternalServerError)
 746                 return
 747         }
 748         if params.delimiter != "" {
 749                 resp.CommonPrefixes = make([]commonPrefix, 0, len(commonPrefixes))
 750                 for prefix := range commonPrefixes {
 751                         resp.CommonPrefixes = append(resp.CommonPrefixes, commonPrefix{prefix})
 752                 }
 753                 sort.Slice(resp.CommonPrefixes, func(i, j int) bool { return resp.CommonPrefixes[i].Prefix < resp.CommonPrefixes[j].Prefix })
 754         }
 755         resp.KeyCount = len(resp.Contents)
 756         var respV1orV2 interface{}
 757
 758         if params.encodingTypeURL {
 759                 // https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html
 760                 // "If you specify the encoding-type request
 761                 // parameter, Amazon S3 includes this element in the
 762                 // response, and returns encoded key name values in
 763                 // the following response elements:
 764                 //
 765                 // Delimiter, Prefix, Key, and StartAfter.
 766                 //
 767                 //      Type: String
 768                 //
 769                 // Valid Values: url"
 770                 //
 771                 // This is somewhat vague but in practice it appears
 772                 // to mean x-www-form-urlencoded as in RFC1866 8.2.1
 773                 // para 1 (encode space as "+") rather than straight
 774                 // percent-encoding as in RFC1738 2.2.  Presumably,
 775                 // the intent is to allow the client to decode XML and
 776                 // then paste the strings directly into another URI
 777                 // query or POST form like "https://host/path?foo=" +
 778                 // foo + "&bar=" + bar.
 779                 resp.EncodingType = "url"
 780                 resp.Delimiter = url.QueryEscape(resp.Delimiter)
 781                 resp.Prefix = url.QueryEscape(resp.Prefix)
 782                 resp.StartAfter = url.QueryEscape(resp.StartAfter)
 783                 for i, ent := range resp.Contents {
 784                         ent.Key = url.QueryEscape(ent.Key)
 785                         resp.Contents[i] = ent
 786                 }
 787                 for i, ent := range resp.CommonPrefixes {
 788                         ent.Prefix = url.QueryEscape(ent.Prefix)
 789                         resp.CommonPrefixes[i] = ent
 790                 }
 791         }
 792
 793         if params.v2 {
 794                 resp.NextContinuationToken = base64.StdEncoding.EncodeToString([]byte(nextMarker))
 795                 respV1orV2 = resp
 796         } else {
 797                 respV1orV2 = listV1Resp{
 798                         CommonPrefixes: resp.CommonPrefixes,
 799                         NextMarker:     nextMarker,
 800                         KeyCount:       resp.KeyCount,
 801                         ListResp: s3.ListResp{
 802                                 IsTruncated: resp.IsTruncated,
 803                                 Name:        bucket,
 804                                 Prefix:      params.prefix,
 805                                 Delimiter:   params.delimiter,
 806                                 Marker:      params.marker,
 807                                 MaxKeys:     params.maxKeys,
 808                                 Contents:    resp.Contents,
 809                         },
 810                 }
 811         }
 812
 813         w.Header().Set("Content-Type", "application/xml")
 814         io.WriteString(w, xml.Header)
 815         if err := xml.NewEncoder(w).Encode(respV1orV2); err != nil {
 816                 ctxlog.FromContext(r.Context()).WithError(err).Error("error writing xml response")
 817         }
 818 }