OpenWrt port — GL.iNet Flint 3 (GL-BE9300, IPQ5332)

Work-in-progress mainline OpenWrt support for the GL.iNet Flint 3 (GL-BE9300) — Qualcomm IPQ5332 (Networking Pro 620, quad Cortex-A53) + 2× QCN6274 Wi‑Fi 7 radios (5 + 6 GHz, on PCIe) + an on-SoC 2.4 GHz radio + RTL8372N 10G switch + RTL8221B 2.5G WAN PHY.

Target: qualcommbe/ipq53xx, kernel 6.12.

This is a target overlay, not a full OpenWrt tree. Drop it into an OpenWrt checkout and build (see below). It carries the board DTS, the kernel patch series, an out-of-tree RTL8372N DSA driver, and the ath12k patches.

Status

Subsystem	State
Boot / procd / SSH	✅ clean procd boot, dropbear up
LAN (RTL8372N via DSA + EDMA/PPE)	✅ 100% ping, 949 Mbit/s iperf3
WAN (eth1, RTL8221B USXGMII)	✅ link up 2.5 Gbps, line rate
VLANs (DSA bridge VLANs)	✅ tagged/untagged validated
Wi‑Fi 5 GHz + 6 GHz (2× QCN6274, PCIe / ath12k)	✅ real clients, SAE + DHCP, simultaneous
Wi‑Fi 2.4 GHz (on-SoC Q6/WCSS / ath12k AHB)	🚧 WIP — driver + DT + firmware done, blocked on a TrustZone PAS step (see below)
eMMC persistence (f2fs overlay, sysupgrade)	✅ config persists across reboots
Standalone cold-boot from eMMC	⚠️ flaky (~50%) — a U-Boot→Linux SDHCI handoff issue (see What's left)

Wi‑Fi 5/6 GHz (solved)

ath12k originally couldn't bind: any PCIe config-space access after boot stalled the AXI bus (config reads worked during enumeration, then hung in pci_read_config_byte(QCN, PCI_INTERRUPT_PIN) inside pci_assign_irq()).

Root cause: the pcie@18000000 node was missing its interconnects (NoC) property. drivers/pci/controller/dwc/pcie-qcom.c (qcom_pcie_icc_init) calls devm_of_icc_get(dev, "cpu-pcie") + icc_set_bw(); with no DT property the CPU↔PCIe fabric path is never given bandwidth, so config TLPs stall once the link goes idle. The 3-line DT fix is patches-6.12/0594-…-add-pcie-interconnects.patch.

With that, ath12k probes both QCN6274 radios and registers two PHYs (non-MLO multi-pdev, per patch 402). Verified on hardware: multiple clients associating end-to-end with SAE + DHCP on 5 GHz (DFS ch100) and 6 GHz simultaneously. The 5 GHz radio is self-managed regulatory and only permits DFS channels here.

Wi‑Fi 2.4 GHz (in progress)

The 2.4 GHz radio is integrated in the IPQ5332 and runs on the on-SoC Q6/WCSS coprocessor (not a PCIe card). The driver side is complete and binds on hardware: the multi-PD remoteproc (qcom_q6v5_mpd, patches 0595–0605), ath12k AHB (ATH12K_AHB + patch 201), the DT nodes (remoteproc@d100000, wifi@c000000, wcss-smp2p, reserved-mem), and firmware staging are all in place. ath12k_ahb recognises the radio (Hardware name: ipq5332 hw1.0) and the Q6 firmware loads.

The blocker: Q6 bring-up fails at the final TrustZone step — qcom_scm_pas_auth_and_reset(pasid=0xD) returns -22 (res[0] = -32). init_image, mem_setup and the segment load all succeed; only the PAS auth is refused. This persists with the correct ath12k firmware (TIP wlan-ap qca-wifi-7/ath12k-firmware/files/IPQ5332 — not the vendor WIFIFW-partition blob, which is for the proprietary cnss driver). GL.iNet and QCA are looking into the TZ side. The nodes ship disabled by default; set remoteproc@d100000 and wifi@c000000 to okay (and stage the firmware) to reproduce.

Discussion: https://forum.openwrt.org/t/gl-inet-flint-3-exploration-gl-be9300-ipq5332/250267.

Deployed

Runs as a live home AP — 5 + 6 GHz with 802.11r (FT‑SAE) + dawn roaming in a multi‑AP network, dual‑VLAN trunk uplink, eMMC f2fs overlay (sysupgrade keeps config).

1 Gbit/s Wi‑Fi to the device; line‑rate to anything behind the uplink. (Standalone cold-boot from eMMC is currently flaky — see below; TFTP/initramfs boot is 100% reliable.)

What's left

• 2.4 GHz on‑SoC radio — blocked on the TrustZone pas_auth_and_reset(-22) above (GL.iNet/QCA engaged). Everything else for it is done.
• Standalone eMMC cold-boot (~50% flaky) — when U‑Boot reads the kernel off eMMC, the SDHCI controller is left in a state Linux's reset only recovers ~half the time (mmc0: Reset 0x1 never completed → mmcblk0 never enumerates → rootwait hangs). The same kernel over TFTP/initramfs (U‑Boot never touches mmc) is 100% reliable. A GCC_SDCC_BCR reset on the &sdhc node helps but isn't deterministic; the robust fix is kernel-side SDHCI re-init. Workaround under test: mmc rescan in U‑Boot before bootm.
• 6 GHz downlink TX — associates and works, but AP→client TX is lossy on some channels; pick a 6 GHz channel with room for the width.
• Upstreaming — the RTL8372N DSA driver especially.

Board notes / gotchas

• Switch port map ≠ silkscreen. On the RTL8372N, port 3 is the CPU port; the four LAN jacks are chip ports 7, 5, 6, 8 (LAN1=7, LAN2=5, LAN3=6, LAN4=8), and port 4 is a 1G secondary CPU/trunk, not a jack. Use the chip port numbers for the DT reg= fields; the silkscreen jack numbers don't match the DSA port names. (The vendor's swconfig tool numbers ports differently again.)
• DSA conduit MTU + vlan‑aware bridge: DSA sizes the CPU conduit for the 8‑byte CPU tag but not the 4‑byte VLAN tag → full‑1500 frames are silently dropped on the CPU/Wi‑Fi path (pings/DNS fine, TCP/SSH stall). Bump the conduit MTU (see the hotplug helper).

Build

git clone https://github.com/openwrt/openwrt.git
cd openwrt
# overlay this repo's target/ and package/ onto the checkout:
rsync -a /path/to/this-repo/target/   target/
rsync -a /path/to/this-repo/package/  package/
./scripts/feeds update -a && ./scripts/feeds install -a
make menuconfig    # Target: Qualcomm Atheros BE -> ipq53xx -> GL.iNet GL-BE9300
make -j$(nproc)

Flash the resulting …glinet_gl-be9300-initramfs-uImage.itb via U‑Boot TFTP, or the …squashfs-sysupgrade.bin.

Note on firmware: proprietary QCA Wi‑Fi firmware is not included. The 5/6 GHz radios use the standard ath12k-firmware-qcn9274 OpenWrt package. To test the 2.4 GHz radio, stage the open-source ath12k IPQ5332 firmware from TIP wlan-ap into /lib/firmware/IPQ5332/.

Layout

• target/linux/qualcommbe/dts/ipq5332-gl-be9300.dts — board DTS
• target/linux/qualcommbe/files-6.12/drivers/net/dsa/rtl8372n.c — RTL8372N DSA driver
• target/linux/qualcommbe/patches-6.12/ — kernel patch series (PCIe/CMN‑PLL/NSSCC, PPE/EDMA/XGMAC, PCS, DSA, the multi‑PD Q6 remoteproc 0595–0605, etc.)
• package/kernel/mac80211/patches/ath12k/ — ath12k patches (dualmac, MLO‑disable, AHB, …)

License

GPL-2.0, matching OpenWrt and the Linux kernel.

Supporting this work

This is a spare-time, single-unit port — the one Flint 3 I own is also the live AP, which makes brick-risky dev and clean A/B testing awkward. If this work has been useful to you, a tip toward a second Flint 3 to develop on is hugely appreciated:

☕ ko-fi.com/perceival

Testing, bug reports, and pointers help just as much — feel free to open an issue.

Credits

Port by Kamil Bienkiewicz (perceivalpercy@gmail.com). Built on upstream OpenWrt and the in-progress mainline IPQ5332 enablement. With help from the OpenWrt community and GL.iNet's engineering team.