benchpi

benchpi, a touchscreen status panel for a Raspberry Pi 5 with a 480x320 ILI9486 SPI LCD hat, running on the bare Linux framebuffer. No X11, no Wayland, no compositor, no browser.

+--------------------------------------------------+
| benchpi                          eth0  192.0.2.17 |
| +------------+ +------------+ +----------------+  |
| |   49.0C    | |    13%     | |     0.03       |  |
| |  CPU TEMP  | |   MEMORY   | |     LOAD       |  |
| +------------+ +------------+ +----------------+  |
| +----------------------------------------------+  |
| | USB                                  up 22m  |  |
| | CH341A programmer                            |  |
| | FT232R USB UART                              |  |
| +----------------------------------------------+  |
| [    REBOOT    ]          [    SHUTDOWN    ]      |
+--------------------------------------------------+

Hardware

Part	Detail
Board	Raspberry Pi 5, 2GB, Debian 13 (trixie), Linux 6.18
Display	/dev/fb0, fb_ili9486, 480x320, RGB565, 960 byte stride
Touch	/dev/input/event1, ADS7846/XPT2046, spi0.1

Touch orientation is corrected in the device tree with swapxy,invy, so the evdev coordinates already line up with the landscape framebuffer.

The user running it needs the video and input groups. The app itself never needs root.

Building

LVGL is not packaged in Debian, so the Makefile fetches the release tarball. It is not checked in.

make lvgl        # download LVGL 9.5.0 into ./lvgl
make -j4         # ~700 translation units, under 30 seconds on a Pi 5

make lvgl has to run first: the source list is a find expanded when the Makefile is parsed.

Running

Take the kernel console off the framebuffer first, or it will corrupt the UI. See "The console owns fb0 too" below for why.

sudo systemctl disable --now gpm      # reads the touchscreen as a mouse
echo 0 | sudo tee /sys/class/vtconsole/vtcon1/bind
./benchpi

echo 1 to that same file gives the console back.

The REBOOT and SHUTDOWN buttons additionally need a sudoers drop-in, or they will display NO SUDO and do nothing:

sed "s/YOURUSER/$USER/" benchpi.sudoers \
    | sudo install -o root -g root -m 0440 /dev/stdin /etc/sudoers.d/benchpi

The sed fills in whoever runs the dashboard. It grants exactly systemctl poweroff and systemctl reboot, and nothing else. It is needed because polkit treats the dashboard's session as remote, having no seat, and demands authentication for org.freedesktop.login1.power-off, which a touchscreen cannot supply. pkcheck says so directly:

$ pkcheck --action-id org.freedesktop.login1.power-off --process $$
Authorization requires authentication and -u wasn't passed.

Running the whole UI as root to avoid that one file would be a much worse trade.

Starting it at boot

sed "s/YOURUSER/$USER/g; s|YOURHOME|$HOME|" benchpi.service \
    | sudo tee /etc/systemd/system/benchpi.service
sudo systemctl enable --now benchpi

The unit runs the UI unprivileged and unbinds the console in an ExecStartPre=+ line, which systemd runs as root regardless of User=. The manual echo 0 > .../vtcon1/bind does not survive a reboot, so without this the LCD comes back up showing a login prompt.

Stopping the service rebinds the console, so the panel becomes a terminal again whenever the dashboard is not running.

Using the CH341A without root

sudo cp 99-ch341a.rules /etc/udev/rules.d/
sudo udevadm control --reload && sudo udevadm trigger

flashrom's ch341a_spi programmer drives the device over libusb, so it needs write access to the /dev/bus/usb node rather than a tty. No kernel driver binds 1a86:5512, so there is nothing to unbind first. flashrom is already installed at /usr/sbin/flashrom, which is not on a non-root PATH here.

What it shows

Hostname, the active IPv4 address and which interface it is on, CPU temperature, memory used, 1 minute load average, uptime, and the product strings of every attached USB device that is not a root hub. Everything comes from sysfs, procfs and getifaddrs(); there are no runtime dependencies beyond libc and LVGL, and nothing is shelled out to.

IPv4 only, and interface-agnostic. A Pi with DHCP reservations on both interfaces still usually has only one of them up, so the panel takes whichever holds an address and prefers the wired one rather than hardcoding eth0 or wlan0. IPv6 is deliberately ignored, because nobody is reading a SLAAC address off a 3.5 inch display.

Labels are only rewritten when their text actually changes. Every redraw is real milliseconds on a 32 MHz SPI panel, so a 1 Hz refresh of unchanged values costs nothing.

REBOOT and SHUTDOWN arm on the first tap and fire on the second, reverting after 3 seconds. A resistive touchscreen on a bench picks up stray contact, and one brush past the panel should not power the machine off.

Idle cost is about 2.8 MB RSS and 0.0% CPU.

Notes

• LVGL 9 talks to the framebuffer through lv_linux_fbdev_create() / lv_linux_fbdev_set_file() and to the touchscreen through lv_evdev_create(). The LVGL 8 display/input driver API is gone; examples written for it will not compile.
• lv_conf.h only lists the settings we change. lv_conf_internal.h supplies a default for every option we leave out.
• fbtft pushes the framebuffer to the panel over SPI via deferred IO. If updates do not appear, try lv_linux_fbdev_set_force_refresh(disp, true), or set LV_LINUX_FBDEV_MMAP 0 to use pwrite() instead of mmap.

The console owns fb0 too

vtcon1 is bound to the same framebuffer, and it will corrupt the UI. What we saw: large rectangles of the screen reverting to pure black, with only the regions LVGL had recently invalidated still showing the right colours.

The cause is that fbtft does not implement fb_blank. When the console blanks the display it therefore falls back to software blanking, which means painting black straight into the framebuffer. /sys/class/graphics/fb0/blank reads 4 (FB_BLANK_POWERDOWN) while the panel is visibly showing a UI. LVGL notices and tries to undo it, which is what this startup warning is:

ioctl(FBIOBLANK): Invalid argument

That is LVGL issuing FB_BLANK_UNBLANK, and fbtft answering EINVAL. It is harmless in itself, but it means LVGL cannot recover the screen on its own. Because the default render mode is LV_DISPLAY_RENDER_MODE_PARTIAL, LVGL only repaints areas it has invalidated, so the blacked-out regions stay black. Note that consoleblank=0 on the kernel command line does not prevent this.

gpm is a second offender. It ships enabled on this image, running gpm -m /dev/input/mice -t exps2, and /dev/input/mice aggregates mouse0, which is the ADS7846. So it reads the touchscreen as a mouse and paints a console cursor onto the framebuffer.

Debugging notes

• The journal is not persistent. /var/log/journal/ exists but is empty, so journald writes to /run and every reboot destroys the evidence. Worth fixing before chasing any crash: sudo mkdir -p /var/log/journal && sudo systemctl restart systemd-journald
• systemd arms the BCM2835 hardware watchdog with a 60 second timeout, so anything that wedges PID 1 is a hard reset with no log at all.
• sudo here is not passwordless. The RPi image sets Defaults timestamp_type=global in /etc/sudoers.d/010_global-tty, so a sudo in any session warms the credential cache for every other session by the same user.
• To screenshot the panel, dump the framebuffer and decode RGB565: cat /dev/fb0 > fb.raw gives 307200 bytes, 480x320 at 2 bytes per pixel with no stride padding. Counting distinct pixel values is a fast corruption check; a clean screen has well under a hundred.
• The display is spi0.0 and the touchscreen is spi0.1, sharing one SPI bus at 32 MHz with fps=30.