Files
NoahandClaude Sonnet 5.5 469f31c278
ci / test (push) Successful in 19s
ci / mcdc (push) Successful in 18s
ci / bench (push) Successful in 5s
ci / qemu-tools (push) Failing after 13s
ci / kernel-symbols (push) Successful in 1m24s
ci / prune-artifacts (push) Skipped
ci / kernel-build (push) Skipped
ci / patches-apply (push) Successful in 17s
ci / quality (push) Successful in 1m2s
ci / badges (push) Skipped
buildroot: nftables cgroupv2 patch reads the handle with the raw syscall (uClibc has no wrapper)
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cxh2piVufTWbzKTmUafTkk
2026-10-09 11:21:17 -06:00
..

luckfox-sdk-patches

The Luckfox SDK (LuckfoxTECH/luckfox-pico, pinned in ../build/vendor.manifest as luckfox-pico) is a ~2 GB vendor checkout and is deliberately not in this repo. Our changes to it were, until now, living only as uncommitted edits in one person's working copy -- which is how a stale second copy on another machine came within one command of silently reverting a day of work. This directory is the tracked form of those changes.

Moved here 2026-09-15 from sdk-patches/ in blueflare/flare-edge (source commit flare-edge@86d4633), so the SDK repo -- not the firmware repo that merely consumes it -- is the single canonical home for every kernel/U-Boot/driver delta. flare-edge's tools/build-firmware.sh and friends now read this directory through the vendor/warden-sdk submodule instead of a same-repo sdk-patches/.

Vendor base: 824b817f8 of https://github.com/LuckfoxTECH/luckfox-pico (Pullrequest 260319 (#355)) -- the same commit ../build/vendor.manifest pins. Every patch here is git diff against that commit and has been verified with git apply --check against a pristine tree at that revision. If the vendor base moves, re-verify before trusting them.

Not patches/, and not sdk/. This repo's top-level patches/ is a different, unrelated thing: the from-scratch Linux 6.18.46 forward-port series (see ../patches/README.md). luckfox-sdk-patches/ patches the legacy vendor tree instead (U-Boot 2017.09 and, optionally, the vendor 5.10.160 kernel) -- naming it luckfox-sdk-patches/ keeps the two unmistakable next to each other. It also isn't named plain sdk/, carrying forward the original naming rule from its flare-edge home: that repo's ui-src/pkgconfig-stub/libdrm.pc resolves ${pcfiledir}/../../sdk, so a directory named sdk/ anywhere it might be checked out alongside would shadow the real vendor SDK checkout. That specific collision cannot happen in this repo, but the name stayed to avoid a second rename and to keep the provenance trail (wiki, commit messages, flare-edge doc citations) legible across the move. Do not rename it.

License

Everything here is a git diff against Rockchip/Luckfox's GPL-2.0 vendor U-Boot (2017.09) and Linux (5.10.160) forks, plus a handful of small standalone files we authored (uboot/include/flare_touch.h, uboot/include/flare_trace.h, the bluez5 wordexp.h compat shim, the mpv Bessel-function patch). The diffs are themselves GPL-2.0 derivative work, consistent with this repo's repo-wide GPL-2.0-only license (../LICENSE); the small standalone files carry no GPL restriction of their own but are distributed here under the same license for simplicity. Unlike ../patches/ (the 6.18 port), no per-driver provenance ledger exists for this directory yet -- everything here predates that convention. If one is wanted, it belongs beside ../kernel/rv1106-enablement/PROVENANCE.md.

Layout

luckfox-sdk-patches/uboot/patches/    diffs against sysdrv/source/uboot/u-boot/
luckfox-sdk-patches/uboot/configs/    new config fragments (not patches -- whole new files)
luckfox-sdk-patches/uboot/include/    new source files
luckfox-sdk-patches/kernel/patches/   diffs against sysdrv/source/kernel/
luckfox-sdk-patches/kernel/configs/   new kernel config fragments
luckfox-sdk-patches/wifi/patches/     diffs against sysdrv/drv_ko/wifi/ (out-of-tree modules)
luckfox-sdk-patches/board/patches/    diffs against project/ and sysdrv/ build files (board .mk,
                                      build drivers, Makefiles, init overlays); may delete files
luckfox-sdk-patches/buildroot/        our delta on the Buildroot tree (see "Buildroot" below)

Paths inside each patch are relative to the vendor SDK root, so apply with -p1 from there:

cd /path/to/vendor/sdk
for p in /path/to/warden-sdk/luckfox-sdk-patches/uboot/patches/*.patch \
         /path/to/warden-sdk/luckfox-sdk-patches/kernel/patches/*.patch \
         /path/to/warden-sdk/luckfox-sdk-patches/board/patches/*.patch; do
    git apply --check "$p" && git apply "$p"
done
cp /path/to/warden-sdk/luckfox-sdk-patches/uboot/configs/*.config  sysdrv/source/uboot/u-boot/configs/
cp /path/to/warden-sdk/luckfox-sdk-patches/uboot/include/*.h       sysdrv/source/uboot/u-boot/include/
cp /path/to/warden-sdk/luckfox-sdk-patches/kernel/configs/*.config sysdrv/source/kernel/arch/arm/configs/

You no longer have to run the above by hand. flare-edge's tools/build-firmware.sh (in the blueflare/flare-edge repo, which carries this SDK as the vendor/warden-sdk submodule) applies everything in luckfox-sdk-patches/ (patches idempotently, config fragments by copy) before it builds -- see its apply_sdk_patches step. The manual recipe is kept here as documentation of what that step does and for one-off use against a pristine tree. Idempotency rule the script uses per patch: if git apply --reverse --check succeeds the patch is already in the tree and is skipped; else if git apply --check succeeds it is applied; else it is a hand- patched SDK carrying an equivalent (or a genuine conflict) and the script warns and continues rather than aborting the build.

Buildroot

The vendor checkout ships buildroot-2023.02.6; WardenOS builds against buildroot-2025.02.8, which is not in the vendor checkout. board/patches/0006-buildroot-2025.patch (project/build.sh) and board/patches/0009-sysdrv-buildroot-2025.patch (sysdrv/Makefile) select it, and ../build/unpack-buildroot.sh puts the pristine tree at sysdrv/source/buildroot/buildroot-2025.02.8: it fetches the tarball with ../build/fetch-buildroot-tarball.sh (sha256-pinned) and leaves an existing tree alone. flare-edge's tools/build-firmware.sh runs it right after applying the patches, then stages buildroot/files/ and applies buildroot/patches/ onto the tree (issue #4).

Completeness check

flare-edge's tools/verify-sdk-patches.sh replays this series onto the vendor base and compares the result with an SDK checkout. It also fails when the SDK has a tracked change (edit or deletion) that no patch, no fragment and no entry in flare-edge's tools/sdk-delta-allowlist.txt accounts for, and when the 86Panel board config names a kernel or U-Boot config fragment that neither the vendor base nor this directory provides. Before that check, five real deltas lived only in copies of one SDK tree (issue #4).

What each patch does

uboot/patches/0001-display-and-boot-timing.patch

The U-Boot boot logo never appeared at all before this, and four separate bugs had to be fixed before it did. Each was measured on hardware, not inferred.

file fix
rockchip_display.c malloc -> calloc. enabled_at_spl is only ever assigned under #ifdef CONFIG_SPL_BUILD, so in U-Boot proper it read uninitialised heap. Garbage read as truthy meant display_enable() skipped the connector pre-enable/enable path -- the only path that reaches backlight_enable().
pwm_backlight.c args.args[2] & 1. The DT has pwms = <&pwm1 0 100000 50000>; assigning 50000 to a bool polarity field inverted the PWM. Measured DUTY=0x26e8 (99.6%) with CTRL=0x13; after the fix CTRL=0x0b. The backlight was running at 0.4%.
clk_rv1106.c Two bugs. rv1106_clk_set_rate() returned a stale ret instead of the computed rate, so it always reported 0 Hz. And DCLK_VOP_DIV_MASK is 5 bits (max ÷32) while DIV_ROUND_UP(1188 MHz, 30 MHz) = 40 wrote 39, truncated to 7 -- the VOP was running at 125 MHz, not 30. Now clamps to 32, picks the closest parent, and retunes CPLL 1000 -> 960 MHz when neither divides evenly, giving exactly 30.000 MHz.
rockchip_panel.c FLARE_PANEL_SETTLE_MS: wait after scanout starts before switching the backlight on. See the comment in the file -- it distinguishes two artefacts that look similar and have different causes.
board.c Release RGB_RES (GPIO0_A1) at the top of board_init() instead of board_late_init(), and FLARE_TRACE instrumentation. setup_download_mode() is deliberately kept first in board_late_init(): moving the logo ahead of it pushed BOOT-key sampling from 1.18 s to ~4-5 s and the key had to be held for eight seconds.
rv1106-u-boot.dtsi u-boot,spl-boot-order = <&emmc>. It was <&sdmmc, &spi_nor, &spi_nand, &emmc> -- the only fitted medium was last. Walking the dead entries cost 267 ms every boot: ~212 ms failing to identify a non-existent SPI-NAND (unknown raw ID 0 0 0) and ~57 ms probing an SD slot this board does not have (sdmmc is wired to the AIC8800 WiFi as SDIO). This file is shared by every rv1106 U-Boot build in the tree; an SD- or SPI-NAND-booting variant would no longer boot.
console.c Timestamp every console line, including from SPL. Needed two fixes: the stock code formats with %5lu.%06lu, which CONFIG_USE_TINY_PRINTF cannot render (it supports only %c %s %u %d %x), and it hooked puts(), which SPL never calls -- tiny printf goes character-at-a-time through putc(). Costs 324 B of SPL.
rockchip_post_csc.c Add #include <common.h>. An upstream SDK bug, not a consequence of our config -- the vendor's own rv1106_defconfig also lacks CONFIG_DM_I2C.

uboot/patches/0002-boot-robustness.patch

Unattended recovery: spl-boot-order.c + spl.c register BOOT_DEVICE_BOOTROM as a last-resort SPL load method so a failed U-Boot load drops to MaskRom instead of hanging, and bootcount/Kconfig exposes BOOTCOUNT_LIMIT so repeated boot failures fall through to altbootcmd. The defconfig change adds the bootcount variables to CONFIG_ENVF_LIST.

Provenance note: this is boot-robustness rather than display work and may overlap with reliability work done in a parallel session. Review before merging.

kernel/patches/0001-panel-mcu-reset-and-logo-route.patch

Two hunks, both display:

  • &panel -- removes reset-gpios and reset-delay-ms. GPIO0_A1 is not the panel's reset. Per the 86Panel schematic, net RGB_RES ties SoC GPIO0_A1, the panel RESET pin (FPC 6) and the NRST of a WCH CH32V003 MCU (U6) together; that MCU's only wired pins are SPI to the same FPC, and its sole job is to clock the panel's initialisation sequence in. Handing the GPIO to panel-simple therefore made drm_panel_prepare() reboot the panel initialiser while the VOP was already scanning out -- which is the faint coloured vertical banding seen for ~1-3 s on every boot, and why adding a "reset" never cured it. Luckfox deleted these same two properties in e2b0ffa22. An earlier revision of our tree re-added them, having read that deletion as an accidental regression. Confirmed fixed on hardware. backlight stays, because U-Boot's rockchip_panel.c takes its backlight handle from this node.
  • &route_rgb -- status = "okay". The rest of the RGB pipeline is already enabled in rv1106-luckfox-pico-86panel-ipc.dtsi; route_rgb was the only piece left off, and without it U-Boot's rockchip_display finds no route and skips drawing entirely.

Deliberately excluded from this patch, though present in the same file in the working tree: the model rename, and a dr_mode = "peripheral" revert that was a bench workaround to keep the USB network gadget stable for flashing. The latter is superseded by the live USB role-switching work on main.

kernel/patches/0005-pstore-ramoops.patch

Adds a ramoops@0f000000 node (1 MiB @ 240 MiB) to reserved-memory in rv1106-luckfox-pico-86panel-ipc.dtsi so pstore/ramoops has a DRAM region that survives a reset. Pairs with the CONFIG_PSTORE* lines in kernel/configs/flare-edge.config: the kernel console ring + oops/panic survive into the next boot (/sys/fs/pstore), which is the ONLY way to diagnose a pre-userspace hang (the c8a3 0x40000 class) with no serial console. Placed HIGH, above the low kernel-image/DTB/initrd zone, so the carve-out can only remove kernel RAM -- it cannot recreate the c8a3 under-reservation collision. Needs a boot.img reflash.

kernel/patches/0006-reboot-mode-maskrom.patch

Adds &reboot_mode { mode-maskrom = <0xEF08A53C>; }; (BOOT_BROM_DOWNLOAD) to the 86Panel ipc.dtsi, so the OS helper warden-maskrom maskrom (reboot(RESTART2,"maskrom")) writes that magic to the boot-mode reg (0xff020200) and the next reset enters true BootROM MaskRom -- where upgrade_tool db+wl / provision-ab.sh flash reliably. The stock mode-loader/-bootloader only reach U-Boot rockusb "Loader" mode, whose multi-partition WL is unreliable (writes report 100% but may not persist). One-shot (PMU-domain reg, cleared by POR). Verified: the DTB compiles and carries mode-maskrom=0xEF08A53C. Needs a boot.img reflash -- and the FIRST deployment still needs one physical-MaskRom flash (chicken-and-egg), after which remote MaskRom flashing works. Independent hunk (end of ipc.dtsi) -- applies in any order vs 0005.

buildroot/files/package/mpv/0002-change-j1.patch

Swaps mpv's j1(x) (video/out/filter_kernels.c, backing the EWA/jinc scalers) for a local BesselJ1(). Needed because this target's uClibc-ng does not provide j1()/j0()/jn() -- confirmed absent from libuClibc-1.0.31.so in the arm-rockchip830-linux-uclibcgnueabihf sysroot, the same kind of gap the bluez5_utils wordexp patch works around for a different function.

The first version of BesselJ1() summed a fixed 11-term power series. jinc() multiplies its argument by M_PI before calling it, and mpv's --scale-radius is user-settable, so any radius reaching roughly x=20 already returns garbage: 11 terms had not converged there, so the value was 3.1e6 instead of the correct 0.0668, growing to 3.3e42 by x=1000. The current version sums a term-to-term-ratio series while each new term is still negligible (stable for the range where that converges) and falls back to the standard second-order asymptotic expansion for Bessel functions beyond that, checked against the host's real j1() across x in [0, 5000]: max absolute error 5e-8. tools/bessel-j1-test.sh pulls whatever BesselJ1() the patch currently carries out of the diff and checks it the same way, so a regression back to an unconverged series is caught there instead of showing up as scaler artifacts on a panel.

buildroot/files/package/bluez5_utils/0001-add-uclibc-wordexp-compat.patch

Adds a wordexp.h at the BlueZ source root: a minimal wordexp(3)/wordfree(3) for uClibc-ng, which this toolchain does not provide. bluetoothctl and a few BlueZ tools include <wordexp.h> unconditionally and only ever pass WRDE_NOCMD ("no command substitution"), so a whitespace/quote tokenizer is a semantically complete substitute -- the same kind of gap the mpv j1() patch above works around for a different function.

WRDE_APPEND lets a caller accumulate words across more than one call onto the same wordexp_t. An allocation failure partway through such a call used to free the words the earlier call had already produced and leave we_wordv pointing at memory realloc() had already relocated or freed -- reachable any time a second WRDE_APPEND call runs out of memory, not only under theoretical low-memory conditions. Fixed by capturing the pre-call word count and only ever freeing/relocating entries at or past it.

tools/wordexp-test.sh extracts wordexp.h from this patch's diff into a scratch dir (nothing else in the tree has a copy -- it is a full new-file patch, applied only during a firmware build) and runs tools/wordexp_test.c against it: plain and quoted tokenizing, an unterminated quote, growth past the initial capacity, WRDE_APPEND across two calls, and the allocation- failure case above. Wired into CI's daemons job.

Seatbelt build changes (docs/firmware-update.md section 6.6)

The firmware-update seatbelt (auto-revert of a bad remote reflash) needs a handful of small U-Boot/OS build changes. Design section 6.6 lists six; items 1, 3, 4 are the low-risk, independently-shippable wiring, and they were -- until now -- present only in one person's hand-patched SDK, which is the exact "stale second copy nearly reverted a day of work" trap that created this directory. They are tracked here and applied by tools/build-firmware.sh, so a clean SDK checkout plus that script reproduces them with no manual editing.

section 6.6 tracked as what it does
item 1 board/patches/0001-86panel-seatbelt-wiring.patch (first hunk) + uboot/configs/flare-bootcount.config Wires the bootcount fragment into the build: appends flare-bootcount.config to RK_UBOOT_DEFCONFIG_FRAGMENT in the 86Panel board .mk. The fragment itself (already tracked, see the Config-fragments table) turns on CONFIG_BOOTCOUNT{,_LIMIT,_ENV} and the extended CONFIG_ENVF_LIST. Without the wiring the fragment is inert and U-Boot never counts boots -- the seatbelt's U-Boot layer (section 6.1) is silently absent.
item 3 board/patches/0001-86panel-seatbelt-wiring.patch (second hunk) + board/patches/0002-build-honor-bootargs-extra.patch panic=10 on the kernel cmdline (sys_bootargs) so a kernel panic reboots after 10 s instead of hanging (section 6.4). The board .mk sets RK_BOOTARGS_EXTRA="panic=10"; the second patch teaches project/build.sh's __GET_BOOTARGS_FROM_BOARD_CFG() to append RK_BOOTARGS_EXTRA verbatim to sys_bootargs (the same place it already appends rk_dma_heap_cma=). That variable flows to RK_KERNEL_CMDLINE_FRAGMENT -> sysdrv/Makefile's KERNEL_CMDLINE_FRAGMENT -> the actual boot args. The hook is generic and inert: a board that never sets RK_BOOTARGS_EXTRA is unaffected.
item 4 kernel/patches/0002-watchdog-enable.patch Enables the hardware watchdog DT node -- &wdt { status = "okay"; }; on the 86Panel board dts. CONFIG_DW_WATCHDOG is already built, but rv1106.dtsi ships wdt disabled, so /dev/watchdog does not exist until this override. Once present, warden-flared pets it only while healthy; a post-boot hang then hard-resets and converges to altbootcmd -> recovery, the backstop to bootcount (section 6.7). Needs a boot.img reflash.

Items 2 (provision recovery env at flash time), 5 (recovery partition + image) and 6 (CONFIG_ENV_OFFSET_REDUND) are not build-time source changes: item 2 is a provisioning-step env write (see tools/envf.py and the flash scripts), item 5 is a partition-table + recovery-FIT change scheduled with a hardware-validation pass, and item 6 is a hardening nice-to-have. They are tracked where they live, not here.

Apply-order note. kernel/patches/0002-watchdog-enable.patch and the existing kernel/patches/0001-panel-mcu-reset-and-logo-route.patch both edit the same board dts, but at non-overlapping anchors (0002 inserts after the &gmac block near the top; 0001 appends &panel/&route_rgb at the tail), so they apply in either order. Verified with git apply --check against the pristine vendor base and by applying both in sequence.

Precise apply targets (relative to the SDK root -- the same -p1 convention as every other patch here):

  • board/patches/0001-86panel-seatbelt-wiring.patch -> project/cfg/BoardConfig_IPC/BoardConfig-EMMC-Buildroot-RV1106_Luckfox_Pico_86Panel-IPC.mk
  • board/patches/0002-build-honor-bootargs-extra.patch -> project/build.sh (function __GET_BOOTARGS_FROM_BOARD_CFG)
  • board/patches/0003-build-fix-rkvar-grep-quote.patch -> project/build.sh (functions unset_env_config_rk, unset_board_config_all)
  • kernel/patches/0002-watchdog-enable.patch -> sysdrv/source/kernel/arch/arm/boot/dts/rv1106g-luckfox-pico-86panel.dts

board/patches/0003-build-fix-rkvar-grep-quote.patch -- required by item 3. unset_board_config_all() reset-sources every RK_* var out of the board .mks with a greedy grep -oh "^export.*RK_.*=", whose .*= runs to the last = on the line. Once item 3 adds export RK_BOOTARGS_EXTRA="panic=10", that grep yields export RK_BOOTARGS_EXTRA="panic= -- an unbalanced quote that makes the sourced temp file a syntax error and aborts every build.sh invocation on a clean SDK (unexpected EOF while looking for matching '"'). The fix narrows both greps (here and unset_env_config_rk) to the variable name only, ^export RK_[A-Za-z0-9_]*=, dropping the value. Without this, a from-scratch checkout cannot build the 86Panel firmware at all; the runner's hand-patched SDK had masked it (the reproducibility gap tools/build-firmware.sh exists to close).

  • uboot/configs/flare-bootcount.config is copied into sysdrv/source/uboot/u-boot/configs/ (it is a whole new file, not a diff); the board .mk patch is what references it by name.

board/patches/0009 to 0013 (issue #4)

Deltas that lived only in the workstation SDK until issue #4. Each patch header explains its change.

patch file what it does
0009-sysdrv-buildroot-2025.patch sysdrv/Makefile BUILDROOT_VER to 2025.02.8 (the tree the buildroot target builds); relaxes the vendor unpack fallback's cp lines
0010-drop-2023-buildroot-package-patches.patch sysdrv/tools/board/buildroot/{busybox,hcitool}_patch/ deletes five 2023-era package patches the fallback would copy into Buildroot
0011-oem-flare-minimize.patch project/cfg/BoardConfig_IPC/luckfox-buildroot-oem-pre.sh flare_minimize: strips camera/ISP files and modules from /oem
0012-drop-s50sshd.patch overlay-luckfox-buildroot-init/etc/init.d/S50sshd deletes the generic sshd starter; dropbear provides SSH
0013-e2fsprogs-build-cflags.patch sysdrv/tools/board/e2fsprogs/Makefile BUILD_CFLAGS so the host helpers build with gcc 14

board/patches/0014-s50usbdevice-unset-role-is-client.patch

0004's S50usbdevice gate read an unset role as router, so a fresh /userdata never composed the USB gadget and the panel had no usb0 until its second boot (flare-edge#278). 0014 makes unset mean client (0), like every other role reader; flare-edge's tests/guards/run-role-default-tests.sh pins the gate to them. It is a separate patch, not an edit to 0004, because build-firmware.sh would skip an edited 0004 as inconclusive on an SDK that already carries 0004 and 0008.

Config fragments

fragment why
uboot/flare-display.config the DRM/panel/PWM symbols the logo path needs
uboot/flare-verbose.config CONFIG_BOOTSTAGE_PRINTF_TIMESTAMP=y -- µs timestamps on every line. ~15 chars/line ~ 1.3 ms/line at 115200, so read instrumented timings as an upper bound
uboot/flare-bootcount.config boot-failure fallback to altbootcmd=download -> MaskRom
uboot/flare-spl-brom.config builds spl_bootrom.o; does not set SPL_ROCKCHIP_BACK_TO_BROM, which would flip SPL_MMC_SUPPORT off
kernel/flare-backlight.config CONFIG_BACKLIGHT_PWM=y. As =m the backlight class device does not exist until userspace insmods it at ~3.3-3.7 s, which is when the logo is being replaced anyway
kernel/flare-wireguard.config CONFIG_WIREGUARD=y for the vendor 5.10 kernel, which builds the recovery image. Named by the 86Panel board .mk; untracked in the workstation SDK until issue #4
kernel/rv1106-usb-host-ethernet.config USB-Ethernet class drivers (with the USB_NET_DRIVERS menu gate) for the vendor 5.10 kernel. Same history as flare-wireguard.config
kernel/flare-fbcon.config # CONFIG_LOGO is not set (the SDK ships the LuckFox fox-head over our logo) plus FRAMEBUFFER_CONSOLE_DEFERRED_TAKEOVER=y (fbcon seizing the fb destroyed the drm_logo image U-Boot handed over)

Measured boot timeline

Instrumented boot, t=0 at the BootROM DDR banner:

milestone before this work after
SPL total (self-reported) 420.6 ms 153.6 ms
U-Boot logo drawn 0.866 s 0.604 s
backlight on (logo visible) 1.157 s 0.695 s
kernel handoff 1.757 s 1.292 s
userspace colour splash 4.335 s 3.87 s

tools/boot-trace.py produces this table from a UART capture. It splits U-Boot from kernel at the Starting kernel line rather than at the kernel banner, because with ignore_loglevel at 115200 the banner is regularly lost to ** N printk messages dropped **.

Still open

  • The white logo starts inverted and fades to correct in its first instant. Believed to be the CH32V003's init still completing when the backlight lands, so we see the tail of the driver IC's internal rail ramp; FLARE_PANEL_SETTLE_MS is currently 250 as a test of that. The cost-free fix, if confirmed, is to release RGB_RES in SPL (~0.03 s) instead of board_init() (~0.21 s) and take the settle back toward zero -- that needs an idblock flash.
  • No display-driver-IC part number has been identified. The panel is a 4-inch 720x720 RGB666 module, flex marked BV-FPC39510A-N-V0. Note the GT911 on this board is the touch controller on i2c3@0x14 and has nothing to do with the display path.