33 — SD-2: SD-card autoboot (separate uboot defconfig, no manual mmc read)

2026-06-21. SD-1 (hand-typed mmc read at the => prompt) → SD-2 (autoboot: the SD card boots to a shell with zero input). Implementation only — board-verify pending. Continuation of 32.

Why a second defconfig

SD-1 needed three hand-typed lines every boot because the single uboot defconfig's CONFIG_BOOTCOMMAND was the NAND mtd read sequence. env=nowhere runs the compiled-default bootcmd every boot, so the fix is a sibling defconfig whose bootcmd is the mmc read sequence — baked in, no prompt interaction.

The NAND defconfig is unchanged; the two defconfigs produce two uboot images (uboot.img / uboot-sd.img), selected by the build/pack/assemble variant.

What changed (5 places, NAND path untouched)

• patch 0005 (patches/uboot/0005-uboot-sd-autoboot-mmc-defconfig.patch): new configs/evb-rk3506_sd_defconfig, identical to evb-rk3506_defconfig EXCEPT the bootcmd:

  CONFIG_BOOTCOMMAND="setenv bootargs 'console=ttyS0,1500000 root=/dev/mmcblk0p3 rootwait rw'; mmc dev 0; mmc read 0x04000000 0x4000 0x5000; bootm 0x04000000"

  root=/dev/mmcblk0p3 — the RKFW SD layout is 3 GPT partitions: uboot=p1 / boot=p2 / rootfs=p3 (NOT p1 like the raw dd image). Load offset 0x4000 + read length 0x5000 match the verified SD-1 manual sequence (boot-sdl-202606211028).
• build-uboot.sh --variant sd: builds the SD defconfig in a throwaway git worktree at the same HEAD (which already has the sd defconfig from 0005), so the SD in-tree build never touches the NAND artifacts in $UBOOT_DIR. Output u-boot-sd-nodtb.bin + u-boot-sd.dtb → $OUT_DIR. tools/mkimage is shared with the NAND build (the defconfigs differ only in bootcmd).
• pack-fit.sh --variant sd: packs ONLY uboot-sd.img — same fit-pack.py Mode A layout as uboot.img, just a different u-boot-nodtb.bin payload. The boot*.img are reused from the default NAND run (media-agnostic).
• assemble-update.sh --sd: uses uboot-sd.img (was uboot.img).
• forge stage_pack_sd: wires build-uboot-sd + pack-fit-sd + pack-sd + assemble-sd; stage_status lists the two new stages.

The Kbuild trap — why a worktree, not make O=

First attempt was out-of-tree make -C uboot O=$OUT_DIR/build-uboot-sd. Kbuild refused:

*** The source tree is not clean, please run 'make ARCH=arm mrproper'
*** in .../third_party/src/uboot

The source tree already has in-tree NAND artifacts (u-boot-nodtb.bin, .o, … from the NAND build), and Kbuild forbids an out-of-tree build over a dirty source tree. mrproper'ing it would destroy the NAND artifacts → break the NAND pack-fit. A git worktree is a pristine separate working tree at the same HEAD → in-tree build works there, and the NAND tree is never touched. (worktree add --detach at HEAD; worktree remove --force in the trap on exit.)

Host verification

• strings: sd uboot bootcmd = mmc read 0x04000000 0x4000 0x5000 + root=/dev/mmcblk0p3; nand uboot = mtd read boot … (unchanged — NAND path confirmed untouched).
• uboot-sd.img 608256 B (fit-pack.py Mode A; uboot loadable sha256 90347127…, fdt c5bd3c40 — same DT as NAND, only the nodtb payload differs).
• update-sd.img 294779466 B (RKFW, chip RK350F, 6 parts, round-trip self-check OK).
• md5 bc572ca0… → /mnt/d/DownloadFromInternet/update-sd-rkfw-autoboot-20260621.img.

Board-verify (2026-06-21, log boot-sdl-2026-06211109) ✅ PASSED — SD-2 closed

RK tool wrote update-sd.img → power on → Hit any key to stop autoboot: 0 (no keypress) → U-Boot 2026.07-rc4-g7e23760a ran the compiled bootcmd → ## Loading kernel from FIT Image at 04000000 + ## Loading fdt → Starting kernel → EXT4-fs (mmcblk0p3): mounted ... r/w + VFS: Mounted root (ext4) on device 179:3 → Run /sbin/init → Welcome to rk-forge buildroot / rk3506 login: root. No => prompt interaction, no UBI — a fully unattended SD boot to a shell. SD-2 closed; the SD path (SD-1 manual + SD-2 auto) is done.