Skip to content

NetBSD 11 on an HP T5325: the XGI Volari Z11 port

The full write-up lives in the gist: hwinfo.md (the hardware survey), patch1.diff (the port), and the boot log. This post is the short version.

The hardware

HP T5325 thin client. Inside it: a Marvell 88F6281 (Kirkwood, “Orion”) running at 1200 MHz — a single ARM926EJ-S rev 1, ARMv5TE, no FPU — with 512 MB of soldered DDR2 (2 × 256 MB on a 16-bit bus), 1 MB of SPI NOR flash, and SATA attached storage. Released around 2011, discontinued long ago, found for nothing.

NetBSD had been on this machine before: a 2014-era NetBSD 6.1.4 build booted to a login prompt on this exact board and this exact U-Boot (MIC 3.4.19-16, Marvell’s fork of U-Boot 1.1.4). That known-good boot is the reference for everything below. The differences between then and now: the machine is running NetBSD 11 instead of 6.1.4, the config was rewritten into an include of the Sheevaplug board plus the deltas this board needs, and there is a framebuffer console.

A brief history of XGI, because it turns out to be one continuous lineage

I grew up with both of these badges — SiS Xabre cards, Trident Super VGA cards — and had no idea they were the same family tree until this port forced the research.

The actual story: SiS was Taiwan’s chipset giant, and its graphics division (the Xabre lineage) was spun off as XGI Technology around 2002, as eXtreme Graphics Innovation, a separate company aimed squarely at competing with ATI and NVIDIA. A few months later XGI absorbed the graphics operations of Trident Microsystems, the 90s Super VGA legend, inheriting the Trident lineage as well. In December 2003 XGI announced the Volari Duo — a card with two GPUs on one PCB, two years before NVIDIA revived 3dfx’s SLI and three before ATI’s CrossFire. Innovative in concept; legendary in execution in mostly the wrong ways, and it never displaced the real competition on the desktop. A couple of years later SiS pulled it back in-house. The Volari Z-series was XGI’s value/embedded line, which is why it ended up in a 2011 thin client with nothing else going for it.

(One more thing I didn’t know: the xgifb driver in this story was formerly an in-tree Linux framebuffer driver. It was dropped from the kernel at some point — I don’t remember which version, and the internet certainly isn’t going to agree on it either — which is why the “debugging” here means reading the dead vendor’s register docs instead of a man page.)

The survey came first

Every number in the hardware notes was read from the running device over SSH, not from datasheets, because the datasheets lie (or at least omit) at exactly the points that matter:

  • SoC reports Feroceon-KW — ARMv5TE, soft-float, so the target is evbarm-earmv5, not -hf.
  • 512 MB RAM, two 256 MB chip-selects, ECC disabled by firmware.
  • 1 MB SPI NOR at 0xf4000000: 512 KB U-Boot, 256 KB SSD firmware, and three 64 KB environment blocks (fw_setenv target lives in one of them).
  • Two SATA ports, only port 0 populated: the on-board 486 MB “SM224” that came with ThinPro on it. NetBSD 11.99.9 is installed on it now, with roughly 48 MB free.
  • Two gigabit ports (mvgbe0/mvgbe1), but note the naming trap: mgi0 is Marvell’s Linux-side name, and NetBSD will not pick up the root device from Marvell’s own rootdev=mgi0 convention. It silently falls through and can pick the wrong disk — back when there was an SP310 on port 1, that meant booting into the wrong system. You pass root= to bootm explicitly (today that’s root=wd0a, the on-board SM224; in 2014 it was root=wd1a), and to hell with convention.
  • Video is the XGI Volari Z11 (PCI 18ca:0027, class 0x030000), BAR0 = 16 MB linear framebuffer at 0xe8000000, BAR1 = 256 KB MMIO, and a legacy VGA register window on PCI I/O space that Marvell’s own driver logs as “XGIfb: Relocate IO address: 1000 [00001030]” — i.e. relocated to BAR2+0x30.
  • Two 16550A serial ports are fitted (MMIO, irq 33/34)… but nobody brought out the UARTs to the case. Video is the only output a human can reach, and there is no way to attach a serial console. That matters a lot when your kernel doesn’t come up.

The build, and the config it’s on now

HPT5325 exists as a kernel set in etc/etc.evbarm/Makefile.inc so distrib/evbarm/installimage can boot it. The config itself is an include of the Sheevaplug config plus the board deltas:

  • makeoptions KERNEL_BASE_PHYS=0x00800000 / KERNEL_BASE_VIRT=0xc0800000 — the kernel loads above the 8 MB boundary that U-Boot and its malloc pool occupy.
  • options MVSOC_CONSOLE_EARLY, VERBOSE_INIT_ARM, CONSPEED=115200 — early output, for when you have a serial.
  • options WSEMUL_VT100.
  • xgiframe* at pci? dev ? function ?
  • wsdisplay* at xgiframe? console ?

A few things in the build were patched around to make it go: the HPT5325 kernel set, a dhcpcd FILESDIR trailing-slash fix (the stray // lands verbatim in the METALOG and nbmtree dies on it), and a bsd.doc.mk fix for SECTION-only documents that produced paths with empty components (same nbmtree failure). Plus one marvell_machdep.c change worth calling out: consinit() now calls ukbd_cnattach() when NUKBD > 0, because without it a USB keyboard can enumerate fine and still never become a console keyboard — wskbd0 at ukbd0 (mux ignored) with no : console keyboard at the end of the line is the tell.

The actual hard part: xgiframe, or dying without a console

The XGI Volari Z11 has no BIOS video POST on ARM — VGA_POST is an x86 thing. There is no firmware to hand you a mode. The generic vga driver attaches on class code alone, so it cannot do anything with the Z11, and the in-tree xgiframe driver (a port of Linux’s old xgifb framebuffer driver — itself since dropped from the kernel) had to learn to do the whole mode set itself. Which is what the patch is. Some of the steps:

Matching the hardware. The PCI ID is known as PCI_PRODUCT_XGI_VOLARI_Z11 and the VGA class code lets it attach, but binding doesn’t mean a usable mode comes out of the other side.

Unlocking the register file. Write 0x86 to SR05 (XGI_IND_SIS_PASSWORD, from Linux’s sis.h — a live chip read-modifies it to 0xa1; the earlier SR01 = 0x12 thing was a misreading of the header names).

Selecting CRT1 and disabling CRT2. The Z11 has no external video bridge, so everything in vb_setmode.c that programmes CRT2 / LVDS / the second CRTC is dead-ended with an early return (pVBInfo->hasVB == HASVB_NONE). Before the mode set, xgiframe_pre_setmode() zeroes CR30 (that’s the CRT1-vs-CRT2 output mux) and sets CR31 = 0x40 | 0x20 (driver mode | VB output disable), plus CR33 = refresh-rate index. This is what makes the chip scan CRT1 — the only CRTC with timing on it — instead of the unprogrammed second CRTC that XGIInitNew() leaves pointed at an LVDS panel this board doesn’t have.

Mode tables. The Z11’s mode table gained the entries the driver was missing (1152×864, 1280×800, 1366×768, 1368×768, 1440×900, 1680×1050, 1920×1080, 1920×1200, 8/16/32bpp) by hand from the datasheet, plus the refresh rates for them. The default mode is 0x47 — 800×600×16 in RGB565, 1600-byte pitch — which is the same DEFAULT_MODE Linux uses.

Instrumentation. Every register write in the mode set now goes through aprint_normal() (CRTC timing, CRT1 tables, pitch, offset, VCLK, output enables), because that is the only observability you get on a machine with no serial port brought out to the case. Plus a write/readback probe on four word offsets spanning the visible area (0x5a5a0000 + i) to settle the question of whether the CPU and the chip are looking at the same DRAM — they are — and to prove the stride is what you think it is.

Keyboard. The ukbd_cnattach() patch above, so a USB keyboard can become the console keyboard and not just enumerate. (Earlier revisions of the config carried options WSDISPLAY_MULTICONS, which mirrors every console character to the serial port as well — the difference between “video may be black, serial is not” and debugging completely blind.)

One detail from the actual boot that the patch doesn’t show: the PCI firmware handed over the BAR0 framebuffer aperture at 0x40000000, outside the 0xe0000000-0xe0ffffff prefetch window the root complex expects, and the driver reprogrammed it to 0xe0000000 itself before mapping the linear framebuffer. That’s the kind of thing you only find out when the hardware writes back.

And then it booted

Booted over TFTP from egiga0 into bootm — this log is the installer, loading its kernel and handing it the USB installer image as the root filesystem. The image identifies itself as NetBSD/hpt5325 11.99.9, built 2026-10-09. This is a 11.99.9 development snapshot, not the 11.0 release. The interesting lines, eliding the arm32_kernel_vm_init memory setup spam:

HP>> tftp
using egiga0 device, our IP 10.25.1.6, server 10.25.1.2
Filename '/netbsd.ub'
Load address: 0x2000000
HP>> bootm 0x2000000
   Image Name:   NetBSD/hpt5325 11.99.9
   Load Address: 00800000
   Entry Point:  00800000
## Transferring control to NetBSD stage-2 loader (at address 00800000) ...
[   1.0000000] NetBSD/evbarm (hpt5325) booting ...
[   1.0000000] NetBSD 11.99.9 (HPT5325) #4: Fri Oct  9 11:20:17 PDT 2026
[   1.0000000]  sq@msi.netcrave.local:/home/hp-netbsd/obj/sys/arch/evbarm/compile/HPT5325
[   1.0000000] total memory = 512 MB
[   1.0000000] cpu0 at mainbus0 core 0: Sheeva 88SV131 rev 1 (ARM9E-S V5TE core)
[   1.0000000] mvsoc0 at mainbus0: Marvell 88F6281 Rev. A0  Kirkwood
[   1.0000000] com0 at mvsoc0 unit 0 offset 0x12000-0x1201f irq 33: ns16550a, 16-byte FIFO
[   1.0000000] com0: console
[   1.0000000] xgiframe0 at pci0 dev 1 function 0: vendor 18ca product 0027 (rev. 0x00)
[   1.0000000] xgiframe0: BAR0 base 0x40000000 outside window 0xe0000000-0xe0ffffff, reprogramming to 0xe0000000
[   1.0000000] xgiframe0: BAR2: offset 0x18 raw 0xf2000001 addr 0xf2000000 size 0x80 (already in window)
[   1.0000000] xgiframe0: XGI Volari: regs 0xf2000000, fb 0xe0000000 (65536 KB)
[   1.0000000] xgiframe0: is_console = yes
[   1.0000000] xgiframe0: display straps: CR48[2:0] = 0x03 (DVI)
[   1.0000000] xgiframe0: register window responding (SR05 = 0xa1)
[   1.0000000] xgiframe0: CR33 = 0x00 (refresh index must be 0)
[   1.0000000] xgiframe0: connector: DVI (CR38 = 0xa0, CR30 = 0x21)
[   1.0000000] xgiframe0: timing and pitch programmed
[   1.0000000] wsdisplay0 at xgiframe0 (kbdmux ignored): console (default, dumb emulation)
[   1.0000000] mvsata0 at mvsoc0 unit 0 offset 0x80000-0x87fff irq 21: Marvell Serial-ATA Host Controller (SATAHC)
[   2.1799692] mvsata0 port 0: device present, speed: 3.0Gb/s
[   2.1799692] wd0 at atabus0 drive 0
[   2.1799692] wd0: <SM224>
[   4.7899750] boot device: <unknown>
[   4.7899750] root device: sd0a
[   5.3710206] root device: sd0a
[   5.4210207] root file system type: ffs
[   5.4410241] WARNING: preposterous TOD clock time
[   5.5030557] WARNING: CHECK AND RESET THE DATE!

NetBSD/evbarm (Amnesiac) (console)

login:

Worth noting what this boot actually was: the installer. The machine was brought up by netbooting the kernel over TFTP from egiga0 and handing bootm the USB installer image (sd0a, a 14 GB PNY Verbatim on the internal USB bus) as the root filesystem to land in sysinst — and from there NetBSD 11.99.9 was installed onto the on-board SM224. That’s why the log shows boot device: <unknown> and root device: sd0a: it is the installer environment, not the finished system.

So the “win” buried in the log is bigger than the login prompt suggests: TFTP-netbooted kernel, framebuffer console, USB installer, sysinst completed, on-board disk now boots NetBSD 11.99.9 — with roughly 48 MB free on a 486 MB drive. NetBSD fits, barely, with nothing to spare. The remaining reminders that this is a porting milestone rather than a deployment: /etc came up read-only in the installer environment, the hostname is Amnesiac (nothing configures one), and the TOD clock is unset (preposterous TOD clock time — CHECK AND RESET THE DATE!).

Between wsdisplay0 at xgiframe0 and xgiframe0: timing and pitch programmed is the whole arc of the work: PCI reprogrammed, register file unlocked, straps read, mode committed, and the chip scanning a framebuffer that the CPU also sees through the aperture.

The summary

This is not a “port of the week” post. It’s a thin client HP shipped with ThinPro Linux on it, on a Kirkwood SoC old enough that NetBSD’s entire ARM story had to be re-established from a 2014 boot log before any of this could start. The port is the sum of: a from-scratch hardware survey done over SSH, a config that is now an include of an existing board instead of a copy, three tiny build-system fixes, a framebuffer console driver taught to do the mode set by hand with no BIOS to hide behind, and a week of aprint_normal() on a machine whose only output device you can’t see until it works — and then, at the end, the login prompt you see in the log above.

If you want to do the same, the gist is the map — including the parts that failed before they worked.