Backup of the mainline-kernel port for the OrangePi 4 Pro (Allwinner
A733 / SUN60IW2). Target: mainline Linux 7.4-rc3+ (666539eecdf4,
2026-11-16), aarch64, NVMe rootfs (Gentoo/OpenRC), serial console on ttyS0.
666539eecdf4. See Rebase to 7.4-rc3 for what changed. It also adds patches 0023–0039: working Ethernet, a Debian SD-card boot, the BL31 memory reservation, the AXP318W PMIC driver, the 26 MHz crystal fix and HDMI (see SD-card boot and Ethernet). Theprebuilt/artifacts and the module paths below still describe the 7.3 build deployed on the board.
| path | what |
|---|---|
kernel/patches/ |
patch series (51) on top of upstream 666539eecdf4 (torvalds/linux), ordered — apply with git am |
rootfs-overlay-debian-sd/ |
Debian 13 SD-card setup for the mainline kernel: U-Boot script and powervr blacklist |
configs/a733-orangepi4pro.config |
working kernel .config used for all board builds |
aic-port/src/ |
AIC8800D80 Wi-Fi and Bluetooth drivers ported to 7.4 |
aic-port/vendor-src/ |
untouched vendor AIC8800 tree |
aic-port/PATCHES.md |
what changed from the vendor AIC8800 sources |
sunxi-drm-port/ |
out-of-tree port of the vendor display stack (DE/TCON/HDMI) to kernel 7.4 — see sunxi-drm-port/NOTES.md for the Kconfig-emulation fix and build/deploy state |
prebuilt/ |
the exact deployed .ko set + Image (md5 da253ed6) + dtbs/ (board dtb md5 de54b62e, boot.scr/boot.cmd) |
rootfs-overlay/etc/ |
Gentoo rootfs config we changed: conf.d/modules (modules="sunxidrm aic8800_fdrv aic8800_btlpm"), modules-load.d/gpu.conf (sunxidrm), init.d/aic8800-bt (BT attach service) |
The Imagination GPU driver (powervr) port lives in its own repository.
git init linux && cd linux
git fetch --depth=1 https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git 666539eecdf4db623f5e3c3e55cf9f2dfad24aa4
git checkout FETCH_HEAD
git am /path/to/kernel/patches/*.patch
cp /path/to/configs/a733-orangepi4pro.config .config
make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- olddefconfig
make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- -j$(nproc) Image dtbs modules
On this base the kernel release is 7.4.0-rc3-g<sha>. The deployed 7.3 build
had vermagic 7.3.0-ge369c7b14bbe-dirty (Image md5 da253ed6), DTB
sun60i-a733-orangepi-4-pro.dtb (md5 de54b62e).
The series was rebased from f47d22362 onto 666539eecdf4 (7.4-rc3+, 2026-11-16).
Verified:
- the 22 patches
git amcleanly onto a fresh666539eecdf4 olddefconfigwithconfigs/a733-orangepi4pro.configkeeps every A733 optionImage dtbs modulesbuild with 0 errors and 0 warnings (aarch64-linux-gnu-gcc 15.2): 1707 modules,Imageabout 44 MB- the rebuilt
sun60i-a733-orangepi-4-pro.dtbdecompiles identical to the deployedprebuilt/dtbs/one
Not verified: booting it. Nothing here has run on a board yet.
What changed:
| Patch | Change |
|---|---|
| 0002 pinctrl | The main A733 PIO driver, its register layout (upstream SUNXI_PINCTRL_NCAT3_REG_LAYOUT, offsets identical) and the shared binding allwinner,sun55i-a523-pinctrl.yaml are upstream now. The patch keeps only the R-PIO driver (allwinner,sun60i-a733-r-pinctrl), with SUNXI_PINCTRL_NEW_REG_LAYOUT renamed to NCAT2_REG_LAYOUT (same bit). |
| 0005 bindings | The private A733 pinctrl binding is dropped. Its fix (main PIO has 10 IRQ banks, matching the driver's irq_banks = 10; upstream required 11) is carried into the shared upstream binding. |
| 0011 bindings | snps,dwmac.yaml keeps both sun50i-h616-emac1 (upstream) and sun60i-a733-gmac0. |
| 0017 debug | init/main.c S8 marker kept after upstream's moved after_dashes block. |
| 0020 grab-bag | Uses upstream's main PIO and devm_kzalloc_obj(s), keeping the a733_pcie_snap trace points and the R-PIO IRQ_CTRL/STATUS access skip (needed so PCIe outbound TLPs keep working). The DWC skip_speed_change guard now wraps the upstream-renamed LINK_WIDTH_SPEED_CTRL/SPEED_CHANGE. |
| 0022 (new) | pcie-sun60i-a733.c follows mainline's renamed DWC port-logic macros (all values identical, mapped by definition). |
Behavior difference to test: the main PIO now uses upstream's io_bias_cfg_variant = BIAS_VOLTAGE_PIO_POW_MODE_SEL and irq_read_needs_mux = true. The old private driver used BIAS_VOLTAGE_PIO_POW_MODE_CTL with a 0x08 control offset. Check IO-voltage-sensitive peripherals (SD, eMMC, Wi-Fi SDIO) on the board.
Config drift: olddefconfig turns CONFIG_DRM_PANEL from y to m and adds new upstream symbols. Nothing A733-specific changed.
Out-of-tree modules: these need work before they load on 7.4.
sunxi-drm-port/builds on 7.4 and drives HDMI (see below).aic-port/src/builds on 7.4 and drives Wi-Fi and Bluetooth (seeaic-port/PATCHES.md).
Tested on an OrangePi 4 Pro booting Debian 13 (Orange Pi OS trixie server
image) from an SD card, with the kernel built from this series. Works: SD
rootfs, serial console, NVMe, Ethernet (100 Mb/s link, limited by the
switch used; DHCP, ping and SSH). HDMI works at 1920x1080@60 with sunxi-drm-port/
(framebuffer console, text and colours checked on a monitor). Gigabit is untested.
Wi-Fi works (tested on a 5 GHz WPA2 network), and Bluetooth scans for devices. USB works at USB 2.0 speed, and the RTC and ADC work. The SPI flash reads reliably at 50 MHz with DMA (it is kept read-only). CPU frequency scaling, PWM, the hardware RNG and the GPU don't work yet.
Run the hardware tests with tests/run-on-board.sh <board-ip>.
To reach the board over a USB cable, see usb-gadget/README.md.
To test kernels without reflashing, see ab-boot/README.md: A/B kernel slots with automatic fallback on the vendor U-Boot.
| Patch | What |
|---|---|
| 0023 dts | GMAC0 resets: stmmaceth = RST_BUS_GMAC0_AXI, ahb = RST_BUS_GMAC0 (vendor BSP roles), so stmmac no longer pulses the bus reset after the glue was programmed |
| 0024 dwmac-sun60i | write the syscfg register after ungating the bus clock and releasing the bus reset (it stayed at reset value 0x8000 = internal EPHY), and enable the phy 25 MHz clock |
| 0025 pinctrl | A733 GPIO_POW_MOD_SEL has a 2-bit field per bank (11 = 1.8 V); the generic code wrote 1 bit with inverted meaning. No effect unless a vcc-pX-supply is set |
| 0026 clk | the bringup clock trace prints at pr_debug, not KERN_EMERG (it flooded the console with SD card-detect polling) |
| 0027 dts | new sun60i-a733-orangepi-4-pro-sd.dts (below) |
| 0028 dtsi | reserve the vendor BL31 secure monitor at 0x48000000 (16 MiB, no-map). Without it Linux can allocate over BL31, and U-Boot can move an initrd onto it (the board then hangs after "Starting kernel") |
| 0029–0033 mfd, regulator | AXP318W PMIC support from the upstream series by Andre Przywara and Jerome Brunet, taken from armbian/build#10712. 0033 is rebased onto 7.4 |
| 0034 clk | pll-ref and the RTC hosc run from a new dcxo clock, set to 26 MHz on this board, so PLL-derived rates are correct |
| 0035 dts | the SD dts uses the AXP318W: SWOUT1 powers the Ethernet PHY and SWOUT2 the SD card |
| 0036 dtsi | the DE, VO-TOP, TCON-TV and HDMI nodes use the /soc bus's one-cell addresses |
| 0037 dtsi | PCIe joins PD_PCIE, so genpd sync_state no longer powers it off under NVMe when the display driver binds |
| 0038 dtsi | the DE output endpoint carries TCON index 4, as in the vendor DT |
| 0039 dtsi | the HDMI PHY uses the real DCXO (26 MHz). With 24 MHz the pixel clock was about 8% fast |
sun60i-a733-orangepi-4-pro-sd.dts (the NVMe board dts is unchanged):
- SD: PF0–PF5 muxed;
uart0no longer claims PF2/PF4 (SD lines);broken-cd(PF6 card-detect caused an RCU stall);vmmcfrom PMICSWOUT2(else "no support for card's volts"). - GPU: disabled. udev autoloads
powervr.koand the board hangs about 8 s into boot. - Ethernet (YT8531C):
snps,clk-csr = <4>(the CSR clock is 200 MHz AHB; the auto choice gave a 12.5 MHz MDC)phy-handle- PHY reset on PH16: the board dts's PA14 doesn't exist, since the A733 has no port A
motorcomm,clk-out-frequency-hz = <125000000>(the GMAC's TX clock comes from the PHY)phy-mode = "rgmii-id"withrx-internal-delay-ps = <1900>andtx-internal-delay-ps = <150>, matching the vendor kernel's PHY registers
- PMIC: AXP318W at
0x36on S-TWI0 (i2c@7083000, PL0/PL1). OnlySWOUT1(dc1sw1,VCC-3V3-PHY, which boot0 leaves off) andSWOUT2(dc1sw2) are described, both always on. The vendor BSP calls this chip AXP8191.
Tested with patches 0029–0035 on an OrangePi 4 Pro running Armbian trixie
(edge): the kernel powers the PHY with no userspace help, and Ethernet gets
a 100 Mb/s link with DHCP and working ping. pll-ref reads 24 MHz, pll-peri0
2400 MHz, the PHY clock 25 MHz and the SD clock 50 MHz. The RTC DCXO_CTRL
frequency select reads 2 (26 MHz).
HDMI needs one more boot script setting. With a monitor attached, the vendor
U-Boot edits the DTB for its own display driver and corrupts this device tree, so
the kernel panics. Set setenv drm_debug 1 in the boot script to skip that edit.
Booting with an initrd (for example Armbian) through the vendor U-Boot
also needs, in the boot script, setenv fdt_high 0xffffffff and
setenv initrd_high 0xffffffff, and a DTB with free space (dtc -p): U-Boot
writes bootargs and the initrd range into /chosen and hangs with
FDT_ERR_NOSPACE otherwise. Armbian edge images for this board work this
way (github.com/ijiki16/build, branch sun60iw2-mainline-edge).
Other findings, not fixed here:
- Schematic OPi 4 Pro v1.3.2 shows the YT8531C with its own crystal and marks the SoC clock path as not fitted, but the vendor dts enables the SoC's 25 MHz fanout (PH15). Patch 0024 enables it too. Whether this board needs that clock is untested.
Patches 0023–0028 and 0034–0039 were written with the help of an AI coding
assistant (Claude). Their commit messages say what was tested, and they carry
no Signed-off-by. Patches 0029–0033 keep their upstream authors and sign-offs.
.github/workflows/nightly.yml runs at 03:30 UTC nightly: pins upstream
666539eecdf4, applies the patch series, cross-builds Image + dtb +
modules, then publishes a rolling nightly release with the artifacts.
Board side: NVMe rootfs is mounted on the (SD-booted) board at
/mnt/nvme-root; modules live in
/mnt/nvme-root/lib/modules/7.3.0-ge369c7b14bbe-dirty/extra/ (real dir
7.3.0-g80c82777a58b-dirty, ge369c7b14bbe is a symlink). UART boot
logs are collected in ~/asdasd (check md5 before reading — stale-log
trap).