Skip to content

Apple's open-source XNU kernel on Android phones: its own loader and drivers, Metal on the Adreno. Runs from RAM; a reboot gives the phone back to Android.

C

docs/driver-review/README.md: the index, what was and was not checked, the decisions left to the owners

shimoku0a297a1f912c
diff --git a/docs/driver-review/README.md b/docs/driver-review/README.md
new file mode 100644
index 0000000..27f177e
--- /dev/null
+++ b/docs/driver-review/README.md
@@ -0,0 +1,57 @@
+# Driver review
+
+Ten read-only audits of the kernel's drivers and of the boot loader, then one pass of fixes. The audits were done by
+reading the code (nothing was run against hardware); each finding has a file and line, a failure scenario and a
+confidence. A fix pass re-read every finding, called some of them not worth fixing, and changed the code only where a
+fix was mechanical and safe (a bound, a wider integer, a lock, an error path). **Nothing here was tested on a phone.**
+What was run: the kernel and the loader build from `drivers/` (via `xnu/apply.sh`), boot in QEMU both from the payload
+device and from the volume (`QEMU_ESP=1`), and the host-side test suites pass (`tools/wpa-test` passes for the
+Wi-Fi supplicant change). The phone-only paths (UFS, GPU, DSP, camera, Wi-Fi bring-up) compile, which is all that was
+checked about them.
+
+| Report | Covers | Findings | After the fix pass |
+|---|---|---|---|
+| [input.md](input.md) | touch, input queue, buttons | 10 | most fixed; firmware verification and the 0666 node left (below) |
+| [gpu-display.md](gpu-display.md) | Adreno, IOSurface, framebuffer | 17 | surface and fence checks fixed; **G1 (no GPU command validation) left** |
+| [camera-haptic-torch.md](camera-haptic-torch.md) | camera, flashlight, vibration | 12 | use-after-free read, torch latch, haptic start fixed |
+| [adsp-battery-cpufreq.md](adsp-battery-cpufreq.md) | audio DSP loader, battery, cpufreq | 18 | DSP-supplied lengths and offsets bounded |
+| [sound.md](sound.md) | speakers, microphone, `/dev/hakios-audio` | 16 | volume clamp, table bounds, guard arithmetic fixed |
+| [wifi-bringup.md](wifi-bringup.md) | Wi-Fi subsystem start | 13 | remote lengths and offsets bounded |
+| [wifi-stack.md](wifi-stack.md) | WPA2 supplicant, 802.11, WLFW | 12 | forged message 1 and last-writer-wins fixed; crypto was found correct |
+| [storage.md](storage.md) | UFS, images, **the write guard** | 12 | guard held against every route tried; timeouts and bounds fixed |
+| [loader.md](loader.md) | USB command parser, payload, volume read | 12 | bounds, busy state and half-read payload fixed |
+| [policy-glue.md](policy-glue.md) | AMFI/MAC stand-ins, device nodes, sysctls, patch hunks | 7 | **nothing changed: decisions for the owners** |
+
+Each report ends with a "Fix status" section: what was fixed and where, and what was left and why.
+
+## Decisions that are not mechanical (open)
+
+These are in the reports; they are listed here because a fix is a choice, not a repair.
+
+1. **Persona entitlement removed from `posix_spawn`** (policy-glue P1, in `xnu/hakios.patch`). Any process can spawn
+ a root child. The reason in the patch comment (nobody carries the entitlement) is out of date, because the
+ code-signing stand-in now reads entitlements from signatures. Restore the check, or document the whole system as
+ single-user.
+2. **`PRIV_NET_PRIVILEGED_SOCKET_DELEGATE` granted to every credential** (P4, patch). Same kind of choice.
+3. **Six device nodes are mode 0666** (P3: touch, audio, microphone, camera, haptic, flashlight): any process can
+ read the microphone and the camera and read or forge the touch stream. They are open because every iOS daemon,
+ whatever its user, has to reach them today; tightening them means choosing which user gets which node, and
+ that needs the iOS user space running to check. The volume ioctl that made the audio node dangerous is fixed;
+ the modes are not changed. Until this is decided, a process on this system is trusted with the sensors.
+4. **GPU command streams are not validated and no `CP_PROTECT` table is written** (gpu-display G1). A client can
+ make the GPU write anywhere its SMMU context reaches. Needs hardware knowledge of the Adreno's protection
+ registers; until then, GPU access is as trusted as root.
+5. **Touch controller firmware is downloaded without a checksum or read-back** (input F6), and its header is the
+ open licensing item `hakios.focaltech_fw.h` (`tools/provenance.allow`, `COMPLIANCE.md`).
+6. **Camera: a user-chosen mode can ask for a very large allocation** (CAM-2) needs an ABI decision (a maximum frame
+ size).
+7. **Wi-Fi: a forged message 1 before message 3 can still fail a join** (wifi-stack W1, rest of it): the proper fix
+ keeps a candidate key set and needs a change to the header the supplicant shares with the 802.11 layer.
+8. **The vsync notification is now sent with the framebuffer's lock held** (input/gpu fix G8). It closes a
+ use-after-release of the notification port; it was not run with the iOS user space. If a deadlock shows up in
+ `IOMobileFramebufferUserClient::vsync`, that is the place to look.
+
+## What the audits did not cover
+
+The modem drivers (left out on purpose), XNU itself, `tools/darwin-write.py`, `var_store.c`, parts of `apfs_lib.c`,
+anything that needs a device on the desk (register values, timings, safe limits).