A game can't pair the virtual pad with its speaker, because both ContainerIDs need one real usb_device #282
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: unom/punktfunk#282
Reference in New Issue
Block a user
Opt-in and default off (
PUNKTFUNK_DUALSENSE_USBIP=1). ⚠ Not yet verified on hardware — see "What still needs on-glass verification".The problem
The Linux pad-audio plane is complete and measured working end-to-end (host sink → capture →
0xD1→ client → a real DS5's voice coils, index-exact, confirmed by hand on 2026-08-17). What still fails is a real game binding to our minted sink. Marvel's Spider-Man Remastered under GE-Proton11-5 enumerates our endpoint correctly and then declines to open it.Root cause — two halves, and both must agree
HID side (
winebus.so):get_container_id_for_usb_udev_devicecallsudev_device_get_parent_with_subsystem_devtype(dev, "usb", "usb_device"). That parent is the only source — no bluetooth path, no synthesis, no env or registry knob. Our pad is auhiddevice whose sysfs chain (/sys/devices/virtual/misc/uhid/…) has no USB ancestor, so the walk fails:Audio side (
winepulse.so, GE-added): the endpoint's ContainerID is derived from the sink's PulseAudiosysfs.pathproplist key (PipeWiredevice.sysfs.path), then the same udev walk with byte-identical GUID packing.pulse_add_deviceskips the call entirely when that key is absent — and a 407 MB+pulselog contains zeroget_container_idlines, proving it was never attempted. Our sinks carry nosysfs.path.Why that matters: the GUID packs
PRODUCT,BUSNUM,DEVNUMandUSEC_INITIALIZEDof the node it resolves to. So the two derivations must land on the same liveusb_device— USB topology for the HID alone would not fix this, and two separate attachments can never match by construction.Wine performs no GUID_NULL comparison anywhere; it just writes
{00000000-…}intoSPDRP_BASE_CONTAINERID. All 18 endpoints registered identically, so a trivial null-match would pair the pad with every endpoint — the rejection is the game's. The log corroborates: the game resolved both DualSense endpoints, readFriendlyName/PhysicalSpeakers, then opened neither, and GE reportsNo live Sony controller mono streams were registered.Second, independent blocker: GE's haptic path enumerates real ALSA cards (
snd_card_next→snd_ctl_pcm_next_device→snd_pcm_opendemanding 48 kHz/S16/4ch). Minted PipeWire nodes are not ALSA cards; the log has zeroChecking ALSA cardlines, so that scan never ran either.pad_sink.rs's own module doc already names both gaps as unsatisfiable by a PipeWire graph.The approach
One composite USB device over
vhci_hcdcarrying its own UAC sound card — the only shape that closes both halves with a single node, since both ContainerIDs then resolve to the sameusb_deviceandsnd-usb-audioproduces a genuine ALSA card for GE's scan.Rejected alternatives: HID-only usbip (fixes half; audio stays GUID_NULL); stamping
device.sysfs.pathonto minted sinks (still needs a real USB node, and leaves the card scan blind);snd-aloop(card name is fixed to "Loopback"); separate HID + audio devices (GUIDs cannot match).Cost is low on the axes that usually kill this:
vhci_hcdis already loaded and in use for the Steam Controller 2, packaging already ships the udev rule granting thepunktfunkgroup write on the sysfsattach, andattach_devicewas already a device-agnostic seam. No new privilege, kernel module or packaging.What's here
52aedb30— vendoredusbip-simlearns alternate settings and isochronous transfers. Two bugs the tests caught: ISO OUTactual_lengthmust report bytes accepted (returning the empty reply's length tells the kernel the device swallowed nothing — audio stalls while looking healthy), and ISO completion must be paced bybInterval, because for an audio endpoint the completion rate is the sample clock. Also fixes a latent panic: only bits 1..0 ofbmAttributesare the transfer type, so real UAC endpoints (0x05/0x09) decoded asNoneand hitunimplemented!().e2b37c60— the pad's 4-interface composite layout, reproduced from anlsusb -vcapture of a real DS5 and pinned by test to the hardware's publishedwTotalLength 0x00E3.74faee31— capture off the ISO endpoint, sink-minting suppressed in that mode, plus apad-usbip-testdevtest.Gate
Docker,
-D warnings: clippy 0, fmt 0; pf-inject 142, usbip-sim 6, punktfunk-host 601 (+22 new). Three of four full runs green — the intermittent failures (sender_delivers_batches,local_summary_reports_a_native_session_as_streaming, a UDPEINTRin untouched code) are pre-existing: pristineorigin/mainfailed 2 of 3 runs in the same environment.What still needs on-glass verification
None of this has run on hardware — attaching a USB device is not read-only and the test boxes were mid-session. One command:
It reports in failure order: device enumerated →
snd-usb-audioproduced the real ALSA card → theusb_devicenode both ContainerId derivations must land on → per-pair peaks (ch0/1 speaker, ch2/3 coils). Then setPUNKTFUNK_DUALSENSE_USBIP=1and re-run the game, expectingCreated container IDin winebus, a non-nullMMDevice_Register Container id, andRouting DualSense.The one genuinely unproven assumption is that
snd-usb-audiobinds cleanly overvhci_hcdand its ISO OUT stream survives the loopback. If it does not, the staged fallback is the HID-only rung: attach the pad without the audio interfaces and stampdevice.sysfs.pathon the minted sinks pointing at an interface node under that sameusb_device— not the device itself, since winepulse starts its walk from the parent and pointing at the device resolves to the root hub and mismatches. That closes the ContainerId half without ISO, leaving GE's card scan blind.Deliberately not included
degrade_if_no_uhidstill gates DualSense on/dev/uhidbefore the transport ladder, which is now over-broad for a pad that does not need it; and the vendored server answers URBs one at a time, so audio pacing can delay a HID poll (it lands near a physical pad's own 4 ms polling, but slower than uhid — fixing it means per-URB tasks, since USB/IP already permits out-of-order completion). No CHANGELOG entry, since the feature is unvalidated and opt-in.A USB Audio Class device needs both and the crate had neither, so the only USB devices it could express were HID ones. Alternate settings: upstream emits exactly one interface descriptor per interface with bAlternateSetting hardcoded to 0. UAC requires alt 0 to be the zero-bandwidth setting and the streaming endpoint to live on alt 1, so the host can release bus bandwidth while the stream is idle. Endpoint descriptors also gained an in-descriptor tail, because a UAC isochronous endpoint is the 9-byte form carrying bRefresh + bSynchAddress rather than the plain 7. Isochronous transfers: the wire parser already read number_of_packets and the iso_packet_descriptor table off the socket, and then dropped both on the floor, always replying with an empty table — which stalls any ISO endpoint. The reply now restates the table per packet. Two details are load-bearing: - actual_length means "bytes the device transferred", and that differs by direction. On IN it is the payload produced; on OUT there is no payload and it is the number of bytes *accepted*. Reporting the empty OUT reply's length would tell the kernel the device swallowed nothing, and an audio stream would make no progress while looking perfectly healthy. - completion is paced by bInterval x packet count, because for an audio endpoint the completion rate *is* the device's sample clock. snd-usb-audio advances its PCM pointer from URB completions and has no other time reference, and vhci_hcd does not throttle the server side (the same reason the interrupt path is already paced). Also fixes a latent decode bug the audio endpoints would have tripped over: only bits 1..0 of bmAttributes are the transfer type — the rest carry the synchronisation and usage type, so a real UAC endpoint reads 0x05 or 0x09. Matching the whole byte yielded None for both and fell through to unimplemented!(), panicking the connection task. Every endpoint the crate shipped with was a plain 0x03 interrupt, so nothing had noticed.