github.com/godbus/dbus/v5 v5.2.4, patched for lys-lib

A copy of github.com/godbus/dbus/v5 v5.2.4 (4e46d5e, BSD 2-Clause, LICENSE; the fd-array bug is unfixed there and on master, which is v5.2.4 as of 2026-10-28), its non-test sources and lys-lib's own tests, carried as packages of the lys-lib module and imported as github.com/stubbedev/lys-lib/third_party/godbus (package dbus), .../godbus/introspect and .../godbus/prop. Every lys process on D-Bus talks to it through this copy, directly or through dbusx.

Why it lives inside the module

As with third_party/oksvg: Go ignores a dependency's replace directives, so a patched godbus wired in through go.mod (lys-shell's third_party/godbus-dbus, which this copy replaces) reached only that module's own builds, and anything importing its packages built against upstream. As packages of the lys-lib module, the patches reach every application importing lys-lib, with no replace.

Patches

Drop the copy once upstream carries every fix below.

• decoder.go, case 'a': an ah decoded with the message's fds made a []UnixFDIndex and appended UnixFD values to it. The reflect panic is a string, which Decode's recover does not turn into an error: up to v5.2.2 every message carrying an fd array (xdg-desktop-portal's Email attachment_fds) arrived with an empty body, and since v5.2.3, which re-panics what is not an error, it crashes the read loop. The slice is now []UnixFD, and an index past the fds is an InvalidMessageError. Pinned by lys_fd_array_test.go.
• encoder.go, case variantType, and lys_encode_as.go: a Variant's value is encoded by the Variant's signature. decode returns a struct as []any, which the Go-type encoder wrote as an array of variants under the struct signature: a malformed message, which dbus-daemon answers by disconnecting the sender. Echoing a decoded GIcon (DynamicLauncher's PrepareInstall) killed the portal's bus connection. Pinned by TestDecodedStructsReencode.
• lys_file_reply.go (with sig.go, encoder.go, dbus.go, export.go): *os.File marshals as h, and the files an exported method replies are closed once the reply is sent. A method returning a bare UnixFD had no safe moment to close its copy, since the reply goes out after it returns. The portal's clipboard pipes and PipeWire remote ride on this. Pinned by TestAFileReplyIsSentThenClosed.
• lys_ordered.go (with conn.go's read loop): DispatchInOrder runs an interface's named methods one at a time in arrival order on one goroutine, through a queue bounded at 256 that holds back the read loop when full; every other call still runs on a goroutine of its own. Upstream dispatches each call on its own goroutine, so calls a sender does not wait on race: xdg-desktop-portal forwards RemoteDesktop input that way, and a click could land before the motion that placed it. Pinned by TestOrderedCallsRunInArrivalOrder.
• prop/prop.go, set: a new value is stored into a fresh allocation that replaces the property's, not into the existing one. Get and GetAll hand out the stored value (a slice shares its backing array) and the reply is encoded after the read lock is released, so storing in place raced with that encode and could rewrite a reply mid-flight; the race detector caught it on a fake NetworkManager's ActiveConnections. Pinned by prop/lys_set_test.go.
• default_handler.go and export.go: the handler's object table is read and changed only under its lock. PathExists, introspectPath and unexport read it unlocked, and export added a path's object and then its interface in two separate steps, so exporting and unexporting from several goroutines (lys-secrets' prompts and sessions) raced, and a call could find an object between the steps. addInterface and deleteInterface now do each change in one locked step. Pinned by TestConcurrentExportsAndCalls.
• Import path: every source imports github.com/stubbedev/lys-lib/third_party/godbus in place of github.com/godbus/dbus/v5, and the copy has no go.mod of its own. Upstream's tests, CI files and fuzz corpus are not carried; the tests named above pin the patches, the bus-dependent ones (lys_bus_test.go) on a private bus from dbustest.

The copy follows lys-lib's Go version, with golangci-lint's modernize fixes applied to the v5.2.2 sources: reflect.TypeFor and TypeAssert, strings.SplitSeq and CutLast, range over ints and over Fields/Methods, slices.Backward and strings.Builder where upstream predates them. No behaviour changes; upstream's tests passed under -race on a private session bus when they were applied, in lys-shell. The v5.2.2..v5.2.4 changes are carried as upstream wrote them.

No dbus-launch

getSessionBusPlatformAddress returns ErrNoSessionBus where upstream runs dbus-launch. That forked a session bus nothing ever stopped, once per call: test runs without a bus left 596 daemons behind, which used up the host's inotify instances, and a program joined a bus no other program was on.