boards: TESTING is dead config, say it with TAGS

README.md

@@ -33,8 +33,10 @@ Rootless cross-compilation of Gentoo stages using crossdev and hakoniwa
 | odroid-xu4 | armv7a | mainline v7.2 | signed BL1/BL2/TZSW + U-Boot | `-O2 -march=armv7ve -mtune=cortex-a15.cortex-a7 -mfpu=neon-vfpv4` | testing |
 | orangepi-5-ultra | aarch64 | mainline v7.2 | TFA + U-Boot + rkbin | `-O3 -mcpu=cortex-a76.cortex-a55+crc+crypto` | testing |
 | pentium-mmx | i586 | mainline v6.18 | BIOS (no firmware) | `-O2 -march=pentium-mmx` | testing |
+| dell-dimension-4100 | i686 | mainline v7.2 | BIOS (no firmware) | `-O2 -march=pentium3` | testing |
 | premier-p550 | riscv64 | mainline v7.2-rc1 | vendor QSPI (OpenSBI + U-Boot) + extlinux | `-O3 -march=rv64gc_zba_zbb` | testing |
 | vsrves01 | riscv32 | mainline v6.18 + patches | VSDSP6/VSOS DDRLoad (no U-Boot/OpenSBI) | `-Os -march=rv32ima_zicsr_zifencei` | experimental (hardware dead) |
+| ch32v467 | riscv32 | mainline v7.1 + patches (buildroot) | none: MCU firmware is the machine (tiny-rv32ima) | `-Os -march=rv32ima_zicsr_zifencei` | experimental (never run on hardware) |
 
 ## CLI
 
@@ -244,7 +246,7 @@ by `uboot`.
 | `BOARD_GCC_VERSION` | no | Pin gcc: `15` (slot), `15.2` (prefix), or exact version (default: highest installed slot) |
 | `KERNEL_TAG` | no | Kernel git ref (default: top-level `TAG`) |
 | `KERNEL_ARCH` | no | Linux `ARCH=` value (default: auto from `BOARD_ARCH`) |
-| `BUILD_STEPS` | no | Build pipeline steps (default: deps checkout bootloader kernel assemble pack); custom step names require a matching `override-<step>.sh` hook |
+| `BUILD_STEPS` | no | Build pipeline steps (default: deps checkout bootloader kernel assemble pack); custom step names require a matching `override-<step>.sh` hook and build with the cross toolchain |
 | `BOOT_PIPELINE` | no | Ordered bootloader stages (default: `("opensbi" "uboot" "syslinux" "grub")`; `()` = none) |
 | `BOOT_EXTLINUX` | no | `true` makes `assemble` write `/extlinux/extlinux.conf` |
 | `BOOT_APPEND` | no | Kernel arguments added to that entry |
@@ -411,7 +413,6 @@ Distinct from `enter` below, which opens a shell in the container the *build*
 runs in: the build host, with the cross compiler, and the target mounted at
 `/target`.
 
-### Debugging a build
 ### Debugging a build
 
 `enter` opens a shell in the very container a build step runs in -- same
@@ -457,7 +458,6 @@ The measurement is the ABI and ISA checks below, against the finished image.
 They are the only thing that can fail a build, because they are the only thing
 that reads the artifact.
 
-### ABI verification
 ### ABI verification
 
 Nothing in a binary package records the libc or the compiler that produced it.

boards/dell-dimension-4100/README.md

@@ -0,0 +1,122 @@
+# Dell Dimension 4100 (Pentium III Coppermine, i815E)
+
+Retail Socket 370 desktop, 2000 vintage.  Pentium III Coppermine at
+866/933/1000 MHz (MMX and SSE, no SSE2), Intel 815E chipset (FW82815 GMCH
+plus FW82801BA ICH2), up to 512 MB PC133, one serial port, two USB 1.1
+ports, Ultra ATA disk, Phoenix BIOS (newest A11).  Dell's OEM build of
+Intel's D815EEA; the UART is on an SMSC LPC47M102 Super I/O.
+
+Nothing here has been booted on the hardware.  Every claim below comes
+from the kernel tree, the chipset datasheets and Dell's archived tech
+specs, not from a serial log.
+
+## Boot chain
+
+No firmware stage of any kind: no U-Boot, no OpenSBI, no vendor blob.
+
+```
+BIOS (Phoenix A11, on-board flash)
+  -> MBR boot code = GRUB boot.img, patched at byte 92 with the LBA of core.img
+  -> core.img in the gap at LBA 1, prefix (hd0,msdos1)/grub
+  -> /grub/grub.cfg on the FAT32 boot partition
+  -> /bzImage
+```
+
+`grub-mkimage` runs in the sandbox; the i386-pc modules come from the
+crossdev prefix, built by the same portage package version as the host
+tool, so the module ABI matches.  `genimage.cfg` writes only the first 440
+bytes of `boot.img` so the disk signature and partition table survive.
+
+## Disk layout
+
+MBR, 512-byte sectors.
+
+| Offset | Size | Contents |
+|---|---|---|
+| 0 | 440 B | GRUB `boot.img` |
+| 440 | 4 B | disk signature, half of the root `PARTUUID` |
+| 512 | up to 1 MiB | GRUB `core.img` |
+| 1 MiB | 128 MiB | boot partition, FAT32, type 0x0c, bootable |
+| 129 MiB | rest | root partition, ext4, type 0x83 |
+
+Everything is inside the first 8.4 GB, so it is reachable whether the BIOS
+answers the INT 13h LBA extension check or falls back to CHS.
+
+Root is named `PARTUUID=${BOOT_DISK_ID}-02`, not `/dev/sda2`: the disk is
+`sda` only while it is master on the primary IDE channel, and this machine
+has two channels and a CD-ROM.  `root=LABEL=` cannot be used at all, with
+or without a device path: `early_lookup_bdev()` in `block/early-lookup.c`
+takes only `PARTUUID=`, `PARTLABEL=`, `/dev/` and a raw device number, and
+`PARTLABEL` is GPT-only.  There is no initramfs to resolve anything else.
+
+## Kernel
+
+Mainline v7.2, `i386_defconfig` plus two fragments.
+
+`kernel-config/pentium3` sets `CONFIG_MPENTIUMIII=y`.  The defconfig names
+no processor family, so the `arch/x86/Kconfig.cpu` choice defaults to
+`M686`; MPENTIUMIII adds `-mtune=pentium3` and turns on
+`X86_INTEL_USERCOPY`.
+
+`kernel-config/dimension-4100` adds `CONFIG_SND_INTEL8X0=y` (8086:2445,
+the ICH2 AC97 controller behind Dell's "SoundMAX 2.0") and
+`CONFIG_VORTEX=y` (the 3C905C-TXM Dell listed as an option), and turns
+`CONFIG_DRM_I915` off.
+
+The defconfig already has ATA_PIIX, BLK_DEV_SD, E100, SERIAL_8250 with
+its console, USB_UHCI_HCD, and EXT4_FS and VFAT_FS built in, which is what
+lets the kernel mount root with no initramfs.
+
+## Compiler flags
+
+`-O2 -march=pentium3 -pipe`, arch `i686`.
+
+Coppermine has MMX and SSE and no SSE2.  `-march=pentium3` is exactly that
+ISA; every `-march` above it turns SSE2 on and an SSE2 instruction faults
+on this CPU.  `-mfpmath` is left at its 387 default on purpose, because
+SSE1 has no double precision.  `i686` rather than `i586` so that Gentoo's
+`default/linux/x86/23.0/i686` profile, whose `CPU_FLAGS_X86="mmx sse"`,
+tells ebuilds the same thing the compiler is emitting.
+
+## Write the image
+
+```sh
+xzcat gentoo-linux-dell-dimension-4100_dev-disk-<timestamp>.img.xz \
+    | sudo dd of=/dev/<target> bs=4M status=progress conv=fsync
+```
+
+Write it to the PATA disk in a USB enclosure or on another machine's IDE
+channel, then put it back as master on the primary channel.  The root
+partition grows to the end of the disk on first boot (`grow-rootfs`).
+
+## Console
+
+Serial is COM1, `ttyS0` at 115200 8N1, and it is the last `console=` on
+the command line, so it owns `/dev/console`.  `console=tty0` comes first,
+so the monitor sees the boot messages too.
+
+## Known not to work
+
+- **Integrated video is unaccelerated VGA text.**  Nothing in this tree
+  binds the 815's GMCH: i915's `pciidlist` starts at `INTEL_I830_IDS`,
+  `INTEL_I815_IDS` (0x1132) is defined in `include/drm/intel/pciids.h` and
+  referenced nowhere under `drivers/`, and the legacy `drm/i810` driver is
+  gone.  `CONFIG_FB_I810` still exists but does not build as shipped:
+  `i810_accel.c` calls `cfb_fillrect`, `cfb_copyarea` and `cfb_imageblit`
+  while its Kconfig entry selects `FB_IOMEM_FOPS`, which pulls in none of
+  them, instead of `FB_IOMEM_HELPERS`, which selects all three.  That looks
+  like an upstream Kconfig bug.
+- **Which NIC is fitted is unknown.**  Both drivers are built in, the
+  integrated Intel 10/100 (`e100`) and the optional 3C905C-TXM
+  (`3c59x`).  One of them is dead weight on any given unit.
+- **ACPI 1.0 and an SMP kernel with LOCAL_APIC/IO_APIC.**  Both are common
+  sources of hangs on 2000-era firmware.  The second GRUB entry boots with
+  `acpi=off noapic nolapic`; if the machine needs it every time, that
+  belongs in the first entry.
+- **The RTC battery is flat.**  `ntpd` is enabled rather than
+  `ntp-client`, which would run before dhcpcd has a lease and give up.
+
+## Default credentials
+
+- root, empty password (development image, change on first login)
+- sshd enabled, `PermitRootLogin yes`, `PermitEmptyPasswords yes`

boards/dell-dimension-4100/board.conf

@@ -0,0 +1,84 @@
+TAGS=("x86" "i686" "pentium3" "i815" "dell" "testing")
+DESCRIPTION="Dell Dimension 4100 (Pentium III Coppermine, i815E + ICH2, BIOS/GRUB)"
+
+# Retail desktop, Socket 370.  Dell's archived tech specs give the
+# configuration: Pentium III at 866/933/1000 MHz, 32 KB L1 and 256 KB
+# on-die L2, Intel 815E chipset, 64 MB upgradeable to 512 MB PC133 SDRAM,
+# one serial port, two USB ports, Ultra ATA EIDE disk, ACPI 1.0, Phoenix
+# BIOS (newest A11).  The board is Dell's OEM build of Intel's D815EEA:
+# FW82815 GMCH plus FW82801BA ICH2, with an SMSC LPC47M102 Super I/O
+# carrying the UART.
+
+BOARD_ARCH="i686"
+CROSS_COMPILE="i686-pc-linux-gnu-"
+KERNEL_ARCH="x86"
+
+# Coppermine is MMX and SSE and no SSE2, which is exactly what gcc means
+# by -march=pentium3: it defines __MMX__, __SSE__ and __FXSR__ and leaves
+# __SSE2__ undefined.  Every -march above it (pentium-m, pentium4) turns
+# SSE2 on, and an SSE2 instruction faults on this CPU.
+#
+# -mfpmath is deliberately unset: the default is 387 and it has to stay
+# 387, because SSE1 has no double precision.
+#
+# -O2, not the -O3 default_cflags() falls back to for an arch it has no
+# entry for.  -O2 is also what Gentoo's own arch/x86/i686 profile ships,
+# so this is the profile's flags with its -march sharpened to the real
+# CPU.  i686 rather than i586 for the same reason: gentoo_profile("i686")
+# selects default/linux/x86/23.0/i686, whose CPU_FLAGS_X86="mmx sse"
+# matches what the compiler is actually emitting.
+BOARD_CFLAGS="-O2 -march=pentium3 -pipe"
+
+# The pin the other mainline boards carry.  The store key is the gcc PVR
+# plus the CFLAGS hash, so leaving it unset moves the cache on every gcc
+# bump.
+BOARD_GCC_VERSION="16.1.0"
+
+# BIOS boot, no firmware stage of any kind: boot.img in the MBR, core.img
+# in the gap before the first partition, grub.cfg on the FAT boot
+# partition.  GRUB_PLATFORMS is load-bearing -- bootloader/grub.rs returns
+# early without it, and it is what puts the i386-pc modules in the
+# crossdev prefix the post-assemble hook copies from.
+GRUB_PLATFORMS="pc"
+# The default pipeline would no-op its way to the same place; naming the
+# one real stage says so out loud.
+BOOT_PIPELINE=("grub")
+
+KERNEL_REPO="https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git"
+KERNEL_TAG="v7.2"
+KERNEL_DEFCONFIG="i386_defconfig"
+# Applied in this order after the defconfig, and every line is asserted
+# afterwards.  pentium3 is the CPU model, dimension-4100 is this machine.
+KERNEL_CONFIG_FRAGMENTS="pentium3 dimension-4100"
+
+BOOT_HOSTNAME="gentoo-dimension-4100"
+
+# rootfs is partition 2, named by the MBR PARTUUID the pack step stamps
+# in.  BOOT_DISK_ID is exported before this file is sourced, genimage.cfg
+# writes the matching disk-signature, and the kernel spells the pair
+# "{sig}-02" itself (block/partitions/msdos.c), so no initramfs is
+# needed.  root=LABEL= is not an option at all: early_lookup_bdev()
+# (block/early-lookup.c) takes only PARTUUID=, PARTLABEL=, /dev/ and a
+# raw device number, and PARTLABEL is GPT-only.  A /dev/ path would be
+# wrong for a different reason: the disk is sda only while it is master
+# on the primary IDE channel, and this machine has two and a CD-ROM.
+BOOT_ROOT_DEV="PARTUUID=${BOOT_DISK_ID}-02"
+
+# The one serial port, COM1 on the LPC47M102 Super I/O.  post-assemble.sh
+# puts console=tty0 in front of this, so boot messages also reach the
+# monitor this machine has; the last console= owns /dev/console.
+BOOT_CONSOLE="ttyS0,115200n8"
+BOOT_SERIAL_TTY="ttyS0"
+BOOT_SERIAL_BAUD="115200"
+
+BOOT_KERNEL_NAME="bzImage"
+
+# ntpd, not ntp-client: the CR2032 in a machine this old is flat, so the
+# RTC is wrong at every boot.  ntp-client is a one-shot that runs before
+# dhcpcd has a lease and gives up; ntpd keeps trying and its default -g
+# steps the first offset however large.  net-misc/ntp and app-admin/metalog
+# are already in defaults/target-packages.txt.
+BOOT_SERVICES=("sshd:default" "dhcpcd:default" "metalog:default" "ntpd:default")
+
+# There is no SD card in this machine; it boots off the PATA disk.
+IMAGE_NAME="gentoo-linux-dell-dimension-4100_dev-disk.img"

boards/dell-dimension-4100/genimage.cfg

@@ -0,0 +1,78 @@
+config {
+    outputpath = .
+    inputpath = .
+    rootpath = gen
+    tmppath = tmp
+}
+
+image rootfs.ext4 {
+    ext4 {
+        use-mke2fs = "true"
+        label = "rootfs"
+    }
+
+    size = 4G
+    mountpoint = "/root"
+}
+
+/* FAT32, not ext4.  Nothing in this machine's firmware reads a filesystem
+ * (the BIOS reads sectors, GRUB reads the rest), so the format is free, and
+ * FAT is the one both grub's fat.mod and a DOS rescue floppy can open. */
+image bootfs.vfat {
+    vfat {
+        extraargs = "-F 32"
+        label = "bootfs"
+    }
+
+    size = 128M
+    mountpoint = "/boot"
+}
+
+image gentoo-linux-dell-dimension-4100_dev-disk.img {
+    hdimage {
+        partition-table-type = "mbr"
+        /* Derived from the board and exported by the pack step; genimage
+         * writes 0 when unset, which every other unsigned disk also claims.
+         * BOOT_ROOT_DEV names the PARTUUID built from this signature, so the
+         * value is load-bearing here, not cosmetic. */
+        disk-signature = "${BOOT_DISK_SIG}"
+    }
+
+    /* 1 MiB in, so LBA 1..2047 stays free for core.img (tens of KiB).  Well
+     * inside the first 8.4 GB, so it is reachable whether the A11 BIOS
+     * answers the INT 13h LBA extension check or falls back to CHS. */
+    partition bootfs {
+        image = "bootfs.vfat"
+        offset = "1M"
+        size = "128M"
+        partition-type = 0x0c
+        bootable = true
+        in-partition-table = "true"
+    }
+
+    /* Last partition; grow-rootfs extends it to the end of the real disk on
+     * first boot with sfdisk ",+". */
+    partition rootfs {
+        image = "rootfs.ext4"
+        offset = "129M"
+        size = ""
+        partition-type = 0x83
+        in-partition-table = "true"
+    }
+
+    /* GRUB i386-pc: boot.img into the MBR boot code, core.img into the gap.
+     * grub-boot.img and grub-core.img are staged by the bootloader step.
+     *
+     * boot.img carries the LBA of core.img at offset 92 (kernel_sector);
+     * grub-install patches it and nothing else does, so it is patched here.
+     *
+     * Only the first 440 bytes of boot.img are written.  Bytes 440..511 are
+     * the disk signature, the partition table and the 0x55AA genimage has
+     * already laid down, and writing 512 bytes here would destroy the
+     * signature root= depends on. */
+    exec-post = 'cp /build/grub-boot.img /build/boot-patched.img && \
+                 printf "\001\000\000\000\000\000\000\000" | \
+                     dd of=/build/boot-patched.img bs=1 seek=92 count=8 conv=notrunc 2>/dev/null && \
+                 dd if=/build/boot-patched.img of=$IMAGEOUTFILE bs=440 count=1 conv=notrunc 2>/dev/null && \
+                 dd if=/build/grub-core.img of=$IMAGEOUTFILE bs=512 seek=1 conv=notrunc 2>/dev/null'
+}

boards/dell-dimension-4100/kernel-config/dimension-4100

@@ -0,0 +1,29 @@
+# Dell Dimension 4100: the peripherals i386_defconfig does not carry.
+#
+# What it already carries was read out of the defconfig, not assumed:
+# ATA_PIIX=y (its PCI table has 8086:244B "Intel ICH2 UDMA 100"),
+# BLK_DEV_SD=y, E100=y (8086:2449 is the ICH2 integrated LAN),
+# SERIAL_8250=y with SERIAL_8250_CONSOLE=y, USB_UHCI_HCD=y, and EXT4_FS=y
+# and VFAT_FS=y both built in, which is what lets the kernel mount root
+# with no initramfs.  VGA_CONSOLE is default y on x86 and not
+# user-selectable without EXPERT, so the text console needs no line here.
+
+# 82801BA ICH2 AC97 controller, 8086:2445 in sound/pci/intel8x0.c, which
+# names it "Intel 82801BA-ICH2".  Dell's solution guide calls the
+# integrated audio "ADI SoundMAX 2.0", an AC-link codec snd-ac97-codec
+# drives.  Not SND_HDA_INTEL, which is what the defconfig carries and
+# which this chipset predates by years.
+CONFIG_SND_INTEL8X0=y
+
+# 3Com 3C905C-TXM, which Dell's tech specs list as an option alongside the
+# integrated Intel 10/100 that CONFIG_E100 already covers.  Only opening
+# the case says which is fitted, so both are built in.
+CONFIG_VORTEX=y
+
+# The 815's integrated video has no driver in this tree that can bind it.
+# i915's pciidlist starts at INTEL_I830_IDS; INTEL_I815_IDS (0x1132) is
+# defined in include/drm/intel/pciids.h and referenced by nothing under
+# drivers/, and the legacy drm/i810 driver is gone.  Dropping it saves
+# megabytes of always-resident text on a 512 MB machine and loses nothing:
+# the console is VGA text mode either way.
+# CONFIG_DRM_I915 is not set

boards/dell-dimension-4100/kernel-config/pentium3

@@ -0,0 +1,12 @@
+# Dell Dimension 4100: the CPU model.
+#
+# i386_defconfig names no processor family, so the choice in
+# arch/x86/Kconfig.cpu takes its own default and lands on M686.
+# MPENTIUMIII is the Coppermine entry, and two things follow from it that
+# M686 does not give: Makefile_32.cpu adds -mtune=pentium3 on top of the
+# same -march=i686, and X86_INTEL_USERCOPY turns on (its depends list in
+# Kconfig.cpu names MPENTIUMIII and not M686).
+#
+# X86_L1_CACHE_SHIFT and X86_USE_PPRO_CHECKSUM come out the same either
+# way, so those are not reasons.
+CONFIG_MPENTIUMIII=y