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