Summary
Cloudreve's WOPI PUT_RELATIVE handler treats X-WOPI-SuggestedTarget as a path, not a filename. It splits the header on / and joins the segments onto the source file's directory with URI.JoinRaw, which feeds Go's url.JoinPath. url.JoinPath resolves ./.. segments, so a slash-bearing target such as a/../../evil.docx collapses to a location outside the source file's directory. The lower-level upload path then validates only the final, already-cleaned basename (evil.docx), which is harmless, and checks ownership against the resolved ancestor — which is still the same user's drive.
A WOPI access token is bound to exactly one file (the route enforces fileId == session.FileID with a 403 otherwise). PUT_RELATIVE escapes that per-file scope: a token issued for one file can create (and, conditionally, overwrite) files elsewhere in the same account.
Root cause (verified at 26b6b10)
1. Token is single-file scoped (the boundary being escaped) — middleware ViewerSessionValidation:
fileId := hashid.FromContext(c)
if fileId != session.FileID { // 403 — token is bound to ONE file
c.Status(http.StatusForbidden); c.Abort(); return
}Route: wopi := noAuth.Group("file/wopi", middleware.HashID(hashid.FileID), middleware.ViewerSessionValidation()); wopi.POST(":id", controllers.ModifyFile) → POST /api/v4/file/wopi/:id?access_token=<token>.
2. PUT_RELATIVE dispatch — routers/controllers/wopi.go:
case wopi.MethodPutRelative: // X-WOPI-Override: PUT_RELATIVE
err = service.PutContent(c, true)3. SuggestedTarget joined as a path — service/explorer/viewer.go:
fileName, _ := wopi.UTF7Decode(c.GetHeader(wopi.SuggestedTargetHeader)) // X-WOPI-SuggestedTarget
fileUriParsed, _ := fs.NewUriFromString(fileUri)
if strings.HasPrefix(fileName, ".") { /* treat as extension */ }
fileUri = fileUriParsed.DirUri().JoinRaw(fileName).String() // <-- path join, not basename
...
subService := FileUpdateService{ Uri: fileUri }
res, err := subService.PutContent(c, lockSession)4. JoinRaw splits on / and normalizes via url.JoinPath — pkg/filemanager/fs/uri.go:
func (u *URI) Join(elem ...string) *URI {
newUrl, _ := url.Parse(u.U.String())
return &URI{U: newUrl.JoinPath(/* PathEscape each elem */ ...)} // JoinPath cleans ./ and ../
}
func (u *URI) JoinRaw(elem string) *URI {
return u.Join(strings.Split(strings.TrimPrefix(elem, Separator), Separator)...)
}PathEscape leaves . unescaped (it is in the unreserved set), so .. segments survive into JoinPath, which resolves them. URI.Name() returns path.Base(path.Clean(path)) — the cleaned basename.
5. Upload checks ownership of the resolved ancestor and validates only the clean basename — pkg/filemanager/fs/dbfs/upload.go:
ancestor, err := f.getFileByPath(ctx, navigator, req.Props.Uri) // URI already traversal-normalized
...
if _, ok := ctx.Value(ByPassOwnerCheckCtxKey{}).(bool); !ok && ancestor.OwnerID() != f.user.ID {
return nil, fs.ErrOwnerOnly // same-user -> passes
}
...
if err := validateNewFile(req.Props.Uri.Name(), req.Props.Size, policy); err != nil { // checks "evil.docx" only
return nil, err
}validateFileName rejects / \ : * ? " < > | and bare ./.. — but the traversal is already gone by the time it sees the basename.
Validation performed
Independent validation against commit 26b6b10 in a clean sandbox.
Source-verified (static): the full chain confirmed verbatim — single-file-scoped token (403 on mismatch) → PUT_RELATIVE dispatch → DirUri().JoinRaw(SuggestedTarget) → url.JoinPath normalization → ancestor ownership check (same-user passes) → basename-only validation of the cleaned name.
Dynamic (control-flow executed): the full binary is not buildable offline here (modules behind an unreachable Go proxy, embedded frontend, DB). I built and ran a harness using the real Go net/url stdlib plus the verbatim Join/JoinRaw/DirUri/Path/Name/PathEscape/shouldEscape and the validateFileName gate, driving the same transformation PUT_RELATIVE performs. Source = cloudreve://my/folder/current.docx:
SuggestedTarget resolved URI final basename validator
"copy.docx" cloudreve://my/folder/copy.docx "copy.docx" ACCEPT
"a/../../evil.docx" cloudreve://my/evil.docx "evil.docx" ACCEPT <- ESCAPED to /
"a/../../../top.docx" cloudreve://my/top.docx "top.docx" ACCEPT <- ESCAPED to /
"sub/evil.docx" cloudreve://my/folder/sub/evil.docx "evil.docx" ACCEPT <- different subdir
".pdf" cloudreve://my/folder/current.pdf "current.pdf" ACCEPT
"a%2f..%2f..%2fenc.docx" cloudreve://my/folder/a%252f..%252f.. "a%2f..%2f..%2f" ACCEPT (NO escape)The headline payload a/../../evil.docx deterministically resolves to cloudreve://my/evil.docx (account root) with a clean, accepted basename. Output matches the original audit probe exactly. Honest caveat: a leading non-.. segment (e.g. a/) is required to prime the join; a single ../evil.docx does not cleanly escape, and URL-encoded separators (%2f) do not traverse through this path (they are re-escaped into one literal segment). Only literal / separators work.
