Dell XPS 13 9315 2-in-1 — camera support on Linux

Reverse-engineering and building working Linux driver support for the cameras on the Dell XPS 13 9315 2-in-1 (Alder Lake, Intel IPU6ep), specifically:

• Rear camera: Samsung S5K3J1, ACPI INT346D — the main focus of this project, and the furthest along. Streams real image data, has a working autofocus lens (VCM) with a from-scratch Linux driver, and runs real continuous contrast-detection autofocus in libcamera's software ISP. Phase-detect autofocus (PDAF) hardware is present but not usable on this kernel — see "Not fixed / known dead ends" below. See STATUS.md for full current state.
• Front camera: OmniVision HI556, ACPI INT3537 — works via mainline hi556.c, no kernel changes needed. Its libcamera image-quality tuning (autoexposure, colour) has had substantial work in this project even though the kernel driver is untouched — see "Building" below.
• IR/Windows Hello camera: OG0VA1B, ACPI OVTI00AB — not yet attempted; groundwork only (see docs/windows-agent-findings-2026-07-13.md). Simpler target than the rear camera (no PDAF/VCM).

Layout

• drivers/ — three git submodules, each a real fork carrying this laptop's patches as normal, diffable commits against exact upstream history (not flat/disconnected copies — see "Building" below for how they fit together):
  • linux-xps9315-2in1/ — a fork of Ubuntu's resolute kernel, branched from the exact tag this project tests against (Ubuntu-7.0.0-29.29). Real in-tree commits, each DMI+HID quirk-gated to this laptop:
    • drivers/media/pci/intel/ipu-bridge.c — an ACPI i2c-instantiation fix for the rear camera's ACPI link, the VCM _CRS resource-index fix (this board has 3 I2C resources per sensor, not the common 2, so the VCM lands at a different index), and a link- frequency fix (512MHz, not 848MHz) against real upstream's own S5K3J1 SAUCE patch, which independently added S5K3J1 support with the same wrong frequency value.
    • drivers/platform/x86/intel/int3472/tps68470_board_data.c — TPS68470 PMIC board data for this laptop's DMI model, fixing two real board-wiring bugs inherited from a copy-pasted "Trial C" starting point (avdd→ANA, dvdd/dovdd→CORE/VSIO, matching this board's actual Surface-Go-derived rail layout), and wiring the VCM's regulator to a real consumer instead of an always_on diagnostic hack.
    • drivers/media/pci/intel/ipu6/ipu6-isys-csi2.c — root cause of the rear camera's total capture failure. Mainline never sets bit 1 of PPI2CSI_CONFIG_PPI_INTF (reg 0x204); Windows sets it on every CSI2 port. With it clear the rear sensor's 4-lane/1024Mbps stream never syncs (black screen); with it set, sustained real video. See STATUS.md's "BREAKTHROUGH" entry for the full diagnosis.
  • ipu6-drivers/ — a fork of Intel's own intel/ipu6-drivers, rebased onto the exact upstream commit this project's s5k3j1.c work is based on, carrying: a real link- frequency bug fix (512MHz, not 848MHz — a genuine D-PHY timing mismatch), PMIC power-sequencing/GPIO-order fixes, a crop fix, and PDAF stream plumbing (kernel side works; the libcamera side is a dead end on this kernel's routing restrictions, see STATUS.md).
  • lc898217/ — new driver, not previously packaged anywhere: the rear camera's VCM (lens actuator) chip, identified via Ghidra analysis of the Windows driver (see docs/vcm-investigation-lc898217.md) and adapted from Vasiliy Doylov's real upstream lc898217xc.c submission, with his copyright preserved. Adds the chip-specific Init calibration handshake this variant needs, and fixes a real cross-platform bug in the upstream driver it's based on (missing i2c_device_id table — silently never binds on any CONFIG_OF=n x86 kernel, this one included).
• scripts/ — install/reload/diagnostic helper scripts for testing the above, plus analysis tools (autofocus sweep/hill-climb, AGC exposure measurement from captured frames).
• docs/ — investigation write-ups, session findings, and scoping docs for larger pieces of work (autofocus architecture, .aiqb calibration-data extraction) written before starting them.
• reference/ — supporting evidence: ACPI table dumps, Windows driver artifacts, live register captures, and the Ghidra reverse-engineering scripts/output used to identify the rear camera's VCM chip.

Building

Two halves are needed for a fully working rear camera with real AF/AGC: the kernel drivers (this repo) and a patched libcamera (published separately, see below). Both steps need linux-headers-$(uname -r) installed first.

1. Kernel drivers

git clone --recurse-submodules https://github.com/mward5/camera-mipi.git
cd camera-mipi
sudo apt install dkms   # if not already installed
bash scripts/install-dkms.sh

This is the recommended path: it registers the drivers with DKMS (Dynamic Kernel Module Support), which rebuilds them automatically on future kernel upgrades (apt upgrade installing a new kernel just works, no manual re-run needed). Prefer a one-shot manual build instead? bash scripts/install-custom-modules.sh does the same builds without registering with DKMS — you'd need to re-run it yourself after every kernel upgrade.

Secure Boot: if it's enabled, kernel modules must be signed to load. DKMS handles this automatically using your system's existing Secure Boot key (mokutil --sb-state checks whether it's on) — if you've already enrolled a MOK for any other out-of-tree module (VirtualBox, acpi-call, etc.), that same key gets reused with nothing further to do. If this is your first time, DKMS generates a new key and install-dkms.sh tells you exactly what to run (mokutil --import ...) and what to expect at the next reboot (a one-time "MOK Manager" enrollment screen — inherent to how Secure Boot works, not something any script can skip). On Secure Boot-disabled systems none of this applies.

Either script: reboot afterward — this ensures a genuinely clean ACPI/fwnode state rather than relying on insmod/rmmod cycling, which silently no-ops some of these fixes (see the scripts' own comments for why). To revert, see each script's own printed instructions at the end of a successful run.

All quirks in drivers/linux-xps9315-2in1/ and drivers/ipu6-drivers/ are gated on an exact DMI model match (and, where relevant, ACPI HID) — safe to build and run this on unrelated hardware; the quirks simply won't activate.

2. libcamera

Published separately at mward5/libcamera (branch xps-9315-2-in-1-cameras). Two ways to get it:

• Prebuilt, attested .debs (recommended): download the latest release, built and attested by GitHub Actions directly from that repo's source — verify provenance with gh attestation verify <file> --owner mward5 before installing, then sudo dpkg -i *.deb.
• Build it yourself: see that repo's own build instructions (standard dpkg-buildpackage/meson package build).

After installing, restart the relevant services (systemctl --user restart pipewire wireplumber, or sudo systemctl restart pipewire wireplumber depending on your setup) to pick up the new libcamera.

Not fixed / known dead ends

Recorded so they aren't re-attempted without new evidence — see STATUS.md for the full detail behind each:

• PDAF — the kernel's v4l2_subdev_routing_validate() forbids the 1-sink-to-2-source routing the sideband capture needs (V4L2_SUBDEV_ROUTING_ONLY_1_TO_1), a hard kernel restriction, not a missing setup step. Continuous autofocus instead uses pure contrast-detection (CDAF) in the software ISP.
• Rear camera's privacy/indicator LED — never lights, because the rear camera hangs off the TPS68470 PMIC path, which has no privacy-LED concept anywhere in mainline (the front camera's discrete INT3472 path does, hence its LED works).
• Dual-monitor flakiness — was suspected to be caused by this project's driver stack; root-caused to an unrelated mutter/Thunderbolt-hotplug bug, filed upstream as mutter#4928.

Start here

Read STATUS.md for current state and what's left to do.