path: root/STATE.md

# Open-IOV Project State

*Last updated: 2026-09-21 (PM)*

## Mission

Run **unmodified NVIDIA guest drivers** in VMs with full GPU functionality
(CUDA, graphics, video, AI/ML) on consumer/datacenter GPUs, without
proprietary GRID licenses — via host-side GSP emulation in
`open-iov-vfio.ko`.

## Current milestone

| Phase | Status |
|-------|--------|
| GSP 570.144 firmware decompilation (132 files, 12 RPC types) | ✅ Complete |
| Refactor 1: GSP emulation engine (per-mdev thread, doorbell, shared mem, MSI-X) | ✅ Complete |
| Refactor 2: PCI config space + BAR + MSI-X emulation | ✅ Complete |
| Refactor 3: All 12 RPC types with emulation fallbacks | ✅ Complete |
| Refactor 4: Portable architecture (fw analyzer, dynamic dispatcher, arch/version layers, YAML configs) | ✅ Complete |
| Refactor 5: Guest driver integration (mdev creation, VFIO registration) | ✅ Unblocked, in test |
| Refactor 6-11: CUDA / graphics / benchmarks / migration / Windows / hardening | ⏳ Pending |

**Right now**: the **guest-probe hang root cause is found and fixed** (VFIO
mmap garbage PFN → KVM EPT livelock — see below) plus a sweep of correctness
fixes (mdev remove leaks, eventfd wiring, GSP shadow binding, RPC livelock).
Guest driver **570.144 stock runfile installed** in the VM. The fixed module
is built and pushed; it loads after the next host reboot (a leaked
`gsp_emu` kthread from the pre-fix remove path pins the old module — rmmod
would oops). Post-reboot: guest `nvidia.ko` probe → `nvidia-smi` → teardown
regression → unit test → CUDA validation (`docs/DEBUGGING.md` §5 checklist).

## Recently solved bugs

| Bug | Root cause | Fix |
|-----|-----------|-----|
| `available_instances` stuck at 0, mdev create fails `-EUSERS` | mdev core's shared `parent->available_instances` counter leaked on crashed removes | `mdev_driver.get_available` callback computes availability from the driver's own device list (kernel/main.c) |
| Module unload hang (`open_iov_vfio` stuck "Unloading") | Legacy `nv_vgpu_gsp_init()` booting physical GSP firmware, conflicting with host nvidia.ko; unload blocked in `cancel_work_sync` | Skip legacy GSP boot, use `gsp_emu_init()`; thread-stop via completion with 5s timeout; `cancel_work_sync` before `destroy_workqueue` |
| mdev remove kernel panic (`vfio_device_group_unregister+0xa`) | NULL deref on VFIO group unregister | Defensive cleanup ordering in `nv_vgpu_mdev_remove()` |
| mdev creation `EINVAL` (-22) | Double IOMMU group handling conflicted with `vfio_register_emulated_iommu_dev()` | Let VFIO create its own group; IOMMUFD emulated callbacks wired |
| `Unknown symbol nvidia_vgpu_vfio_probe` | NVIDIA OpenRM built without `KBUILD_EXTRA_SYMBOLS` | Build order automated in `stack_manager.sh` (kernel → nvidia → userspace) |
| `vfio 1: No such file or directory` | VFIO device registration is NOT automatic for mdev | `vfio_register_emulated_iommu_dev()` in probe, unregister in remove (Intel kvmgt pattern) |
| Guest-visible PCI config space wrong | BAR types/offsets, subsystem ID offsets, MSI-X table size | Config space rewritten to match physical TU104 (BAR0 16MB 32-bit non-prefetch, BAR1 256MB 64-bit prefetch VRAM, BAR3 32MB; 32 MSI-X vectors; subsystem IDs at 0x2C/0x2E) |
| **Guest driver probe hard-hangs whole VM (580 & 570.144)** | VFIO `mmap` used `virt_to_phys()` on `vzalloc`'d BAR0 shadow + ioremap'd BAR2/3 → garbage PFN → KVM `make_spte` reserved-bit WARN → EPT-violation livelock (diagnosed via QEMU GDB stub: one vCPU frozen on a single MMIO instruction) | BAR0 mapped per-page via `vmalloc_to_pfn()`; BAR2/3/4/5 map `bars[N].gpa` directly; `nv_vgpu_mmap_virt_to_pfn()` helper |
| **GSP emulation thread never saw guest RPCs** | Engine polled the *host physical* BAR0 doorbell + a local `dma_alloc_wc` buffer; guest writes land in the vzalloc `bar0_shadow` | `gsp_emu_init` binds exclusively to `bar0_shadow` (doorbell/mailbox/PBA aliases); also stopped PBA writes corrupting the host GPU's physical BAR0+0x3000 |
| **Guest never received MSI-X completions** | `SET_IRQS DATA_EVENTFD` handler discarded eventfds (`TODO` + `kfree`); `msix_trigger[]` all NULL; inject used an init-time (stale) copy | Full eventfd wiring: `eventfd_ctx_fdget` per vector, `-1` clears, close releases refs; inject looks up `vgpu->irqs[0].trigger` live |
| **BAR_TRAP/MIGRATE RPC livelock** | dispatch → `gsp_rpc_submit_raw` rings the same doorbell the thread polls → self-triggered infinite loop + 2–30 s stalls | emulation-mode guards return `-ENOTSUPP` with `pr_warn_ratelimited` |
| **mdev remove leaked everything** | remove never `list_del`'d (available_instances stuck), never ran cleanup (leaked `gsp_emu` kthread), never unregistered VFIO (leaked `/dev/vfio/N` + iommu group), never put pdev/freed struct | kvmgt-style remove: wait fds → `vfio_unregister_group_dev` → cleanup work → `list_del` → `pci_dev_put` → `vfio_put_device` |
| `dma_map_page` IOVA corruption | 64-bit `dma_addr_t` stored in `int ret` (truncates ≥4 GB IOVAs under IOMMU) | dedicated `dma_addr_t iova` in `gsp_rpc_alloc_emul` |
| config-space write returned `-1` (`SIZE_MAX`) | `while(count--)` wrap + bogus return expression; function was dead code while the live path did unfiltered byte copies | return fixed; VFIO write handler now routes through the per-register filter (guest can't rewrite vendor ID/BARs) |
| Misleading `MODULE_SOFTDEP("pre: nvidia")` | module has zero nvidia imports (it *exports* `nvidia_vgpu_vfio_*`) | removed; load order is open-iov-vfio **before** nvidia.ko |
| Kernel unit tests didn't build | written against a removed struct-based RPC API | rewritten to drive `gsp_emu_process_rpc()` with crafted mailbox messages; `depends: open-iov-vfio` |

## Known open bugs / limitations

- Guest `nvidia.ko` end-to-end probe not yet validated (fixed module loads
  post-reboot; checklist in `docs/DEBUGGING.md` §5)
- **Memory-model gap**: guest VRAM (BAR1) maps directly to physical BAR1 gpa,
  but `gsp_rpc_alloc_emul` services ALLOC with host system pages — the two
  are not unified; must be resolved before CUDA/DMA-heavy workloads
  (`docs/GSP_EMULATION_ARCHITECTURE.md` §3)
- BAR_TRAP / MIGRATE RPCs fail fast with `-ENOTSUPP` in emulation mode
  (guest's normal BAR access is direct MMIO; real handlers TODO)
- `gsp_emu_process_rpc()` doesn't bump stats counters (only the kthread
  path does) — cosmetic
- WPR callback registration lacks kernel eventfd support
  (`TODO` in driver/wpr — legacy guest-driver tree)
- Migration WPR state save/restore incomplete
- BAR `force_access` path not fully wired (unused-function warning)
- Kernel `SCHEDULE` emulation returns success without real work submission
  (userspace emulation works)
- Async `NOTIFY` events not emulated in kernel (no eventfd/poll yet)
- Single GPU validated so far: TU104 (RTX 2080 SUPER Mobile), GSP 570.144,
  kernel 6.8.0-139-generic
- Live migration protocol implemented (4-phase) but not yet validated
- Windows guest not yet validated (automation scaffolding present)
- MSI-X emulation has dual stores (guest-visible `bar0_shadow` table vs
  `bar_emul_ctx.msix_table[]` used by the currently-unused trap path) —
  documented in `docs/GSP_EMULATION_ARCHITECTURE.md`; harmless with the
  direct-mmap design

## Reproducibility notes

- NVIDIA base: `open-gpu-kernel-modules` tag **610.57.04** + patch series
  in `legacy-patches/nvidia/` (re-brands version string to 570.144 to match the
  GSP firmware ABI; adds the vGPU VFIO interface headers + symbol exports).
- GSP firmware: `nvidia/tu102/gsp/gsp-570.144.bin` from
  [linux-firmware.git](https://github.com/NVIDIA/linux-firmware)
  (pin `2237b99e`), `.fwimage` section extracted with `objcopy`,
  MD5 `13ce12a392a31eeb79a2e6ebd3a4a93a`. `install.sh` does this.
- Firmware blobs, decompiled output, NVIDIA tree, VM images, Ghidra DBs and
  local VM wrappers are intentionally **not** tracked in git.
- Fresh-system bring-up: `./install.sh` (see README.md).

## Hardware/software reference environment

- Host GPU: NVIDIA RTX 2080 SUPER Mobile (TU104, PCI `10de:1ed3`)
- Host OS: Ubuntu, kernel 6.8.0-139-generic
- Guest OS: Ubuntu 24.04.5 LTS (QEMU/KVM, q35, OVMF)
- GSP firmware: 570.144 (RV64IMAC, base `0x04000000`)