xhci98

This project xHCI98 is a WDM generic USB host controller driver for xHCI host controllers targeting Windows 98 SE, ME, 2000 SP4 and 32-bit XP. Although xHCI Controllers offer USB 3.0, this driver runs USB 2.0 on the controller only.

This driver is developed based on Intel's xHCI specification and tested only on Intel machines so far. No guarantees have been made on xHCI implementations from other vendors.

This project is from a solo human with AI-assistance only so bugs are not unexpected. Feel free to report them if you encounter any issues.

This is my 2020 ThinkPad P14s Gen 1 (Comet Lake xHCI, no EHCI) on Windows 98 SE. Connected devices are 7-port hub, USB Ethernet, USB Audio, a USB-to-SATA bridge, two flash drives and a mouse all using the xHCI controller.

Demo video: https://www.youtube.com/watch?v=AU77f9CSbYc

Blog post of this project: https://yeokhengmeng.com/2026/08/xhci98-usb-host-driver/

Motivation

Modern PCs from the mid-2010s onwards use xHCI-only USB chipsets. Windows 98 shipped with UHCI/OHCI drivers (USB 1.1) and gained EHCI (USB 2.0) through the back-ported stack in NUSB. Windows 2000 gained EHCI natively in SP4.

Neither has native xHCI support at all so even basic devices like keyboards, mice and flash drives do not work out of the box without BIOS support. Even with BIOS support, those devices cannot be hotplugged and other devices like Ethernet and Audio are not supported.

This project attempts to fill that gap.

Why USB 2.0 only, on a USB 3.0 controller

The existing usbport.sys this driver depends on does not support USB 3.0 and neither does anything above it. This project driver xhci98.sys is only the miniport underneath that stack.

SuperSpeed would mean rewriting the entire USB HCD (Host Controller Driver) for both operating systems. This is significantly more work than this driver for a speed that most machines running Windows 98 or Windows 2000 is unlikely to effectively use. The xHCI programming guide summarises what it would take.

Every USB 3.x connector (USB4/Thunderbolt included) also carries the USB 2.0 wires and xHCI exposes them as a separate logical port per connector. This driver manages those USB 2.0 ports and leaves the USB 3.x ones unpowered so a SuperSpeed-capable device falls back on the USB 2.0 port and runs at High-Speed.

Installation Steps

The driver needs a USB 2.0 stack (usbport.sys + usbhub20.sys) on the machine first:

• Windows 98 SE: Either install NUSB 3.3 or 3.6 or the SweetLow's stack. For SweetLow's stack, unzip, right-click the USB2.INF at its root then install. Reboot if requested after installing the USB 2.0 stack.
• Windows ME: Use SweetLow's stack only.
• Windows 2000 SP4 and XP SP3 (32-bit): Nothing to install, both OSes at their service pack level already have the stack.

On an xHCI-only Windows 98 SE or ME machine, have the Windows installation CD at hand or the contents on disk as the driver needs some files from there.

Optional but recommended: boot real DOS (not a DOS box inside Windows) and run XHCIQUAL from the XHCIQUAL\ folder. A controller reporting no legacy interrupt pin cannot be driven on either system and there is no software workaround, so find out before you install anything.

Basic read-only quick scan answers in one line. The xhci run below it exercises the controller (handoff, reset, DMA, interrupt, port resets) and ends with a verdict:

Running the full check.

Controller qualified verdict.

XHCIQUAL demo video: https://www.youtube.com/watch?v=Tv6blmBS6Do

Install

1. Put the unzipped package somewhere the machine can read: a floppy, a CD, a shared folder. release\ is the one to install. debug\ is the same driver built for troubleshooting, only install if asked.
2. In Device Manager, find the unrecognised xHCI controller. It sits unclaimed with a yellow mark, usually under "Other devices" such as "Universal Serial Bus Controller".
3. Properties -> Driver -> Update Driver -> Specify a location/Have Disk -> the release\ directory.
4. It installs as "USB 2.0 eXtensible Host Controller (xhci98)" with a "USB Root Hub" underneath it, and neither should carry a warning mark.
5. Reboot if requested.

On Windows 98 with NUSB, do not disable, remove or upgrade this driver in Device Manager. Each of those blue-screens that system. The fault is in NUSB's usbport.sys, the Windows 2000 build of the USB 2.0 stack, not this driver. Microsoft's own USB drivers do the same thing on the same machine. The readme has the way round it for NUSB systems.

The same driver on the same machine survives all three under SweetLow's build of that stack.

What is tested, and what is not

Windows 98 SE is validated on real hardware. Windows 2000 SP4, Windows ME and Windows XP have only ever run in QEMU virtual machines.

Machine	Controller
ThinkPad E460 (2016)	Intel Skylake, Sunrise Point-LP (100-series) PCH. xHCI 1.0, 8086:9D2F, 18 ports (12 USB 2.0 managed, 6 USB 3.0 left unpowered), no EHCI.
ThinkPad P14s Gen 1 (2020)	Intel Comet Lake PCH-LP (400-series). xHCI 1.1, 8086:02ED, 18 ports (12 USB 2.0 managed, 6 USB 3.1 left unpowered), no EHCI.

	State
Windows 98 SE	Validated on real hardware and in VMs. HID, mass storage, USB Ethernet and USB Audio have all run on real xHCI silicon, at a root port and behind hubs.
Windows 2000 SP4	Virtual machines only, including an SMP guest and Driver Verifier. It has never run on real hardware.
Windows ME	One virtual machine only, under SweetLow's USB 2.0 stack (the only stack it is supported with): the driver loads and starts, and a HID mouse, a USB mass-storage device and a composite audio device bind (2026-09-02). Never run on real hardware.
32-bit Windows XP	One virtual machine only (XP Professional SP3): the package installs on an xHCI-only machine with no prompt, the driver loads and starts under XP's own USB stack, and a HID mouse, a USB mass-storage device and a composite audio device bind; disable, enable, remove and rescan in Device Manager all survive. Never run on real hardware.
Intel 7/8-series (XUSB2PR mux), AMD	Never run on either. Everything said about the XUSB2PR port mux comes from Intel's datasheet and Linux, not silicon. The driver does not touch it.
Resume from standby (Windows 2000)	Never executed anywhere. No available VM offers a resumable power transition, and there is no Windows 2000 machine.
Low Speed, USB Audio, hub topologies	Work on Windows 98 hardware in the configurations tried. Not covered: an audio device with bInterval > 1, a USB 1.1 hub under a multi-TT hub, and the Windows 2000 side on silicon.

The devices checked so far, all on the E460 under Windows 98 SE. Each is characterised in test-equipment.md.

Device	VID:PID	Speed	Result
Terminus 7-port hub, multi-TT	1A40:0201	High	The hub every "behind a hub" result below was taken on.
Terminus 4-port hub, single-TT	1A40:0101	High	Characterised. The swap partner for the hub above.
Genesys 7-port hub (two cascaded chips), single-TT	05E3:0608	High	Characterised. A second-tier hub position.
Logitech USB Optical Mouse	046D:C077	Low	Works at a root port and behind the multi-TT hub.
Microsoft Wired Keyboard 600 (composite, two HID interfaces)	045E:0750	Low	Types, once Windows 98's own usbhub.sys is present; since 1.0.0.1 the install has Windows copy it from the Windows 98 CD or CABs.
SanDisk U3 Titanium flash drive	0781:5408	High	Works at a root port and behind the hub.
MSSU10-128GSR and SanDisk 3.2Gen1 USB 3.0 flash drives	090C:2320, 0781:55AB	High (SuperSpeed falls back)	Enumerate at High-Speed on the USB 2.0 port. A file round trip passed.
StoreJet Transcend USB-to-SATA bridge (ASMedia)	174C:5106	High (SuperSpeed falls back)	A drive letter, and a file written and read back with matching contents. The first real bridge chip this driver has done verified I/O through.
ASIX AX88772A USB Ethernet	0B95:7720	High	DHCP lease and traffic on Windows 98 hardware. Also validated on Windows 2000 in a VM. Wedges the machine in three fast replug cycles (the defect below).
Sound Blaster Play! 2 (UAC 1.0 composite)	041E:323D	Full	Plays clean at a root port and behind the multi-TT hub.
Sound Blaster Play! 3, C-Media USB Audio Device (UAC 1.0)	041E:324D, 0D8C:0014	Full	Enumerate and are named by the wizard. Found the Full-Speed bMaxPacketSize0 bug.
Sound Blaster X4 (UAC 2.0, bInterval 3 and 4)	041E:3278	High	Enumerates but does not bind on Windows 98 (one HID devnode at Code 10, no composite parent), so its bInterval > 1 endpoints were never exercised.

ATTO Disk Benchmark on the P14s against the MSSU10-128GSR flash drive trasferring around 18 MB/s read and write from 32 KB transfers upward. The USB 3.0 drive runs at USB 2.0 speed on this driver.

Known limitations

The important limitations. The full list is under "Known limitations" in release-notes.md.

Limitation	Detail
Disabling, uninstalling or upgrading an NUSB driver crashes the machine	A defect in NUSB's usbport.sys which cannot stop a running controller. Rename the existing XHCI98.SYS, reboot, then remove it.
Every device on a root port is reported as High Speed	Reporting the true speed of a slower device crashes usbport as there is no companion controller. A mouse or keyboard on a root port therefore polls at 1, 2 or 4 ms only. If this is an issue for you, put your lower-speed device behind a hub to allow the true speed to be reported.
DisableSelectiveSuspend = 1 is written machine-wide	A suspended xHCI controller cannot see a newly plugged device. A driver uninstall does not remove the value.
Fast, repeated plug and unplug can freeze Windows 98	About twice a second sustained. Ordinary use is fine.

Toolchain and building

The driver is C (C89/C90, no C++ or CRT), built and verified on Windows 11 x64. The toolchain unpacks inside the repository (tools/, git-ignored) and installs nothing to C:\. Every script finds it relative to its own location, so a clone builds wherever it is unpacked.

1. Download the two toolchain archives into tools\ under these exact names:

   File	Source
   tools\MSVC600.zip	MSVC 6.0 - itsmattkc/MSVC600
   tools\WIN2KDDK.EXE	Windows 2000 DDK - KunYi/WDK_DDKArchive

2. Unpack them in place:

   powershell -ExecutionPolicy Bypass -File scripts\setup-msvc6.ps1        # -> tools\MSVC600
   powershell -ExecutionPolicy Bypass -File scripts\install-w2kddk-cabs.ps1 # -> tools\ntddk

3. Build the driver. both (the default) builds release and debug. all adds the emulator-only flavour.

   scripts\build-driver.cmd both
   

   Flavour	Output
   release	src\objfre\i386\xhci98.sys - the shipping driver
   debug	src\objchk\i386\xhci98.sys - the diagnostic build, also shipped
   qemu	src\objchk_qemu\i386\xhci98.sys - emulator-only, never published

4. Build the two tools that ship beside the driver. xhciqual\build.cmd produces XHCIQUAL.EXE, the DOS qualifier, and needs Open Watcom 2.0 at C:\WATCOM (or wherever WATCOM points). That is the only tool installed normally on the host, and the driver never uses it. xhcisnap\build.cmd produces XHCISNAP.EXE, the snapshot reader, with the in-repo MSVC 6.0.

5. Make install media. A .sys on its own is not install media; the INF travels with it, and since 1.0.0.1 nothing else does, because the INF has Windows supply its own usbd.sys and usbhub.sys (since 1.0.1.0, usbport.sys on Windows 2000 and XP; since 1.0.2.0, usbui.dll on all four). For a Windows 98 SE target, first download NUSB 3.3 (nusb33e.exe) from philscomputerlab.com to tools\nusb33e.exe.

   Script	Output
   scripts\package\make-package.ps1	out\pkg-<flavour>\ - media a VM or a machine can be pointed at
   scripts\package\make-release.ps1	releases\<version>\ and out\xhci98-<version>.zip - the published cut

6. Test in QEMU with the qemu-xhci device: a Win98 SE guest, a Win2000 SP4 guest, and an SMP Win2000 guest for race detection.

The version lives in src/xhci_version.h. The INF's DriverVer is a literal that the build's INF gate checks against it. See docs/contributing/build-and-test.md for the VM setup, versioning, packaging and the rest of the procedure.

Repository Layout

src/            Driver source (C), the INF, and the DDK build files
test/           Host-side unit tests for the DDK-free core (test\run-host-tests.cmd)
scripts/        Setup helpers, build wrapper, the import/INF/packaging gates,
                and vm-matrix/ - the automated device matrix
xhciqual/       The DOS hardware-qualification tool (Open Watcom) and its
                bare-metal run logs under results/
xhcisnap/       The snapshot reader for the driver's log channel (both targets)
docs/           Documentation, indexed by docs/README.md
  using/        The user-facing release notes
  contributing/ Roadmap, architecture, build/test/runbooks, diagnostics,
                 invariants, provenance, and numbered design records
  usb-xhci-info/ xHCI programming/data structures, USBPORT ABI, WDM constraints,
                 and controller quirks
  references/   Fetch metadata for external specifications (downloads git-ignored)
releases/       The releases that have been cut, one directory per version
images/         Photographs of the driver on real hardware, used by this README
tools/          The build toolchain, unpacked in place (git-ignored)
external/       Local mirrors of the reference sources (git-ignored)
vm/             Guest images and per-run evidence (git-ignored)
out/            Staged install media from the packager (git-ignored)
.github/        Issue forms - the bug report and the hardware report
AGENTS.md       Guide for AI agents working on this project

Documentation

Everything is under docs/, indexed by docs/README.md.

Using the driver:

• Release notes - what the driver does, does not, and does not yet claim, per target
• Release acceptance test - the fixed procedure for checking a release on a new machine

Working on the driver:

• Build and test - toolchain setup, builds, VMs, install, debugging, packaging, recovery
• Roadmap - project status: what each phase was for, its status, and the two acts left to the owner (the upload, the hand-run acceptance)
• Architecture and implementation invariants
• Source files - what every file in src/ is for
• Failure diagnosis and measured lessons - read these before theorising about a failure
• Design records - the numbered design decisions
• Test equipment, as measured - every hub and device used for hardware validation
• Legal and provenance record

The hardware and the Windows USB stack:

• xHCI programming and xHCI data structures
• USBPORT miniport interface and ABI - the undocumented contract this driver plugs into
• Windows 98/2000 WDM constraints

AGENTS.md is the guide for AI agents, and doubles as the short orientation for a human.

AI usage

This project was developed with substantial help from AI coding agents, Claude Code (Claude Fable 5 and Claude Opus 5) and OpenAI Codex (GPT 5.6 Sol). The agents wrote and revised code, tests and documentation under the guidance in AGENTS.md.

Kheng Meng directed the work, ran the hardware validation on real machines and reviewed what went in.

References

• xHCI specification (Intel), citations verified against revision 1.2c
• USB-IF xHCI backwards compatibility testing (v1.7)
• ReactOS usbport - primary reference for the usbport.sys miniport interface (undocumented by Microsoft)
• Linux xhci-pci.c - controller quirk table
• Haiku XHCI driver and FreeBSD xhci.c - second opinions on hardware details

Neither PDF is tracked here. Fetch your own copies into the git-ignored docs/references/. docs/references/README.md records the versions and SHA-256 sums. Mirrors of the source references can be fetched into external/. See external/README.md.

Licensing and provenance

This project's own source is licensed under the GNU General Public License, version 2 (LICENSE), GPL-2.0-only.

The repository tracks no third-party binary on its own, although the two tool executables it tracks under releases/ carry statically linked third-party runtimes.

• xhci98.sys links no third-party object, runtime or extender.

• XHCIQUAL.EXE embeds the Open Watcom runtime and the DOS/32A extender.

• XHCISNAP.EXE embeds the MSVC 6.0 runtime.

Each ships with a NOTICE.TXT recording it, and the LICENSE scope note says that those runtimes carry their own terms and are outside the grant. Neither quotes the terms themselves.

The full inventory, provenance methods and redistribution boundaries are in docs/contributing/legal-provenance.md, which states facts, not legal conclusions.