sandwich/vfio-native
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: kvm and qemu patches grouped by the detection each clears

Browse Source
This commit is contained in:
Sandwich
2026-09-06 00:14:56 +02:00
commit 2e5269995a
17 changed files with 1912 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

86
patches/README.md Normal file
View File

@@ -0,0 +1,86 @@
# Patchsets, by the detection each one clears
Every patch here is grouped by the VMAware check it targets, so a patchset can be taken or left
on its own. Apply from the relevant source root with `patch -p1 < <file>`.
## KVM - `kvm/`
Built against kernel 7.2.x. Builds as `arch/x86/kvm` only, about five minutes; no full kernel
build is needed. See `packaging/` for a PKGBUILD that does it and keeps it rebuilt.
| Patch | Clears | What it does |
| --- | --- | --- |
| `0001-KVM-SVM-intercept-GP-when-guest-EFER.SVME-is-clear` | `SVM_EXCEPTIONS` | Closes a `FIXME` KVM already carries: a CPL3 SVM instruction faults `#GP` where hardware gives `#UD`. |
| `0002-KVM-x86-emulator-UD-not-GP-for-VMCALL-at-CPL-0` | `KVM_INTERCEPTION` | Adds `PrivUD` so `VMCALL` at CPL>0 raises `#UD` like hardware. |
| `0003-KVM-x86-UD-for-KVM-hypercalls-issued-at-CPL-0` | `KVM_INTERCEPTION` | A CPL>0 hypercall raised no exception at all; now raises `#UD`. |
| `0004-KVM-SVM-intercept-ICEBP-and-skip-it-before-injecting-its-DB` | `DBVM` | SVM re-injected the `ICEBP` `#DB` with RIP still on the `F1` byte; VMX and hardware report the next instruction. |
| `0005-KVM-selftests-verify-ICEBP-DB-reports-RIP-past-the-ICEBP` | - | The selftest for 0004: fails on unpatched SVM, passes with it. |
| `EXPERIMENTAL-0006-runtime-cpuid-passthrough` | `TIMER` | Drops CPUID and RDPRU interception on a 1:1-pinned vCPU and reprograms the AMD Processor Name String MSRs so raw `CPUID` still returns the declared brand. Opt-in via the `cpuid_passthrough` module parameter. |
**0002 and 0003 are a pair.** Either alone leaves `KVM_INTERCEPTION` standing, because the check
tries two stubs and only reports the first one that misbehaves.
The first five are written to be acceptable upstream: they make KVM match the architecture rather
than add anything on top of it, each carries its rationale in the commit message, and all four
pass `checkpatch.pl --strict`. `0000-cover-letter.txt` is the series cover letter. They
have been tested on AMD only; 0002 and 0003 are vendor-neutral code and the Intel reasoning is in
the cover letter.
**`EXPERIMENTAL-0006` is built into the `vfio-native-kvm-dkms` package but off by default.** It
removes the cost the `TIMER` instruction-latency detector measures, and with it the guest reads
raw host CPUID, so it applies itself only to vCPU threads confined to exactly one CPU and must be
enabled through the `cpuid_passthrough` module parameter after the guest has booted.
`docs/ANALYSIS.md` has the measurements and the constraints.
## QEMU - `qemu/`
Against QEMU v11.1.1. Apply in numeric order.
| Patch | Clears |
| --- | --- |
| `01-firmware` | `FIRMWARE` - ACPI link-device names, `PNP0A06` stubs, FADT C-state latencies |
| `02-disk-identity` | `VIRTUAL_DISK`, `DISK_SERIAL` - `QEMU HARDDISK`, empty serials, SCSI inquiry strings |
| `03-pci-ids` | `PCI_VENDOR`, `DEVICES` - Red Hat/QEMU vendor and subsystem IDs, ICH9 layout |
| `04-fw-cfg` | `QEMU_FW_CFG` - the fw_cfg device signature |
| `05-usb-hid` | `QEMU_USB` - USB HID descriptor strings |
| `06-audio` | `DEVICES` - HDA codec and controller identity |
| `07-edid` | none. Carried for other detectors; VMAware does not read EDID. |
| `08-cpu-misc` | none. CPU feature plumbing and one upstream backport. |
`patch -p1 < qemu/01-firmware.patch` and so on, then build with the prefix set to `/opt` so the
system QEMU package is left alone:
```sh
./configure --prefix=/opt/qemu-native --target-list=x86_64-softmmu
ninja -C build && ninja -C build install
```
## Approaches tried and rejected - kept so they are not re-derived
| Approach | Result |
| --- | --- |
| Patched OVMF firmware (`edk2`) | Builds, then hangs before console init at 100% CPU, cause undiagnosed. Stock OVMF is fine and costs only the boot-logo warning. |
| TSC compensation for `TIMER` (BetterTiming) | No detection change, 6x slower boot. |
| CPUID leaf override (Hypervisor-Phantom) | Halved the timing signature, still detected, and *added* a `SINGLE_STEP` detection. |
| Clearing `INTERCEPT_CPUID` outright | Guest never boots. |
| Intercept clearing, minimal hypercall | Triple fault, or no effect. |
`TIMER` has two detectors that OR together. The exception-latency one passes on stock KVM once it
is measured the way VMAware measures it; the instruction-latency one is removable with
`EXPERIMENTAL-0006`. `docs/ANALYSIS.md` has the measurements, including a direct reproduction of
both in `bench/timerprobe.c`.
## Identity values in these patches
Several patches carry a concrete make and model because no real machine leaves the field blank.
Every one of them is a placeholder. Change them before you use this, or a detector gains a
fingerprint shared by everyone who ran the same patch series.
| Where | Value | Replace with |
| --- | --- | --- |
| `02-disk-identity` | disk serial `S0123456789ABCD` | set `<serial>` on the disk in the domain XML; this fallback then never runs |
| `02-disk-identity` | disk and optical drive model strings | a drive you actually own, from `lsblk -o NAME,MODEL` |
| `07-edid` | monitor vendor and model | your own panel, from `edid-decode < /sys/class/drm/*/edid` |
The SMBIOS values are not patched into QEMU at all - they are domain XML, taken from a real
machine with `dmidecode`. Change every serial there too.