Re: [virtio-dev] Re: [RFC PATCH v3 3/3] SPI: Add virtio SPI driver. - Correction

From: Haixu Cui
Date: Wed Mar 06 2024 - 02:48:32 EST


Hello Harald,

In current driver, spi_new_device is used to instantiate the virtio SPI device, spidevX.Y is created manually, this way seems not flexible enough. Besides it's not easy to set the spi_board_info properly.

Viresh Kumar has standardized the device tree node format for virtio-i2c and virtio-gpio:


https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/devicetree/bindings/i2c/i2c-virtio.yaml

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/devicetree/bindings/gpio/gpio-virtio.yaml

In this way, the driver is unified, board customization only depends on the device-tree node. It's easy to bring up spidev automatically.

Look forward to your opinions. Thanks a lot.

Haixu Cui


On 3/6/2024 1:54 AM, Harald Mommer wrote:
Hello,

looked again at my tinny setup and I've to add a small correction.

It's not the way that I've no udev at all there. What is in place there is busybox mdev.

Relevant part of /etc/init.d/rcS:

#!/bin/sh
mount -t proc none /proc
mount -t sysfs none /sys
depmod
modprobe spi-virtio
mdev -s
mdev -d

If I kill the "mdev -d" process my small script below does not make the /dev/spidev0.0 device node appear any more. Of course not, there must be some user mode process which does the job in the device directory.

Regards
Harald Mommer

On 05.03.24 11:57, Harald Mommer wrote:
Hello,

I took next/stable as base giving the exact tag/sha of the current next/stable so that it's known what was used as base version even when next/stable moves. The ordinary next tags are currently not of best quality, gets better, therefore next/stable now. We were on v6.8-rc7 yesterday with next/stable.

VMM is qemu for the driver you have. But it's a specially modified qemu which allows that we use our proprietary virtio SPI device as backend.

Proprietary virtio SPI device is started first, this is an own user process in our architecture. Subsequently the special internal qemu version is started. The virtio SPI driver is compiled as a module and inserted manually by a startup script by "modprobe spi-virtio". The driver goes live immediately.

In this simple setup I do not have udev rules (no service supporting udev => no rules) so no /dev/spidevX.Y automatically after the driver went live. What I'm using to test the latest driver before sending it to the mailing lists is really a naked kernel + a busybox running in a ramdisk. The udev rule I've sent are used on some more complete setup on real hardware.

So without udev I have to bring this device up manually:

In /etc/spidev-up.sh there is a script tp bring up /dev/spidev0.0 manually:

#!/bin/sh
SPIDEV=spi0.0
echo spidev > /sys/bus/spi/devices/$SPIDEV/driver_override
echo $SPIDEV > /sys/bus/spi/drivers/spidev/bind

Afterwards there is /dev/spidev0.0.

In linux/tools/spi there are spidev_test.c and spidev_fdx.c. Those (somewhat hacked locally, and I mean "hacked" to be able to test somewhat more) are used to play around with /dev/spidev0.0.

I can do this on my Laptop which has no underlying SPI hardware which could be used as a backend for the virtio SPI device. The proprietary virtio SPI device has a test mode to support this. Using this test mode the driver does not communicate with a real backend SPI device but does an internal simulation. For example, if I do a half duplex read it always gives back the sequence 01 02 03 ...

For full duplex it gives back what has been read but with letter case changed, in loopback mode it gives back exactly what was sent. With this test mode I could develop a driver and parts of the device (device - real backend communication to an actual SPI device) on a board which had no user space /dev/spiX.Y available which could have served as backend for the virtio SPI device on the host.

Slightly different module version is tested on real hardware with the virtio SPI device not in test mode. "Tested on hardware" means that device + module work for our special use case (some hardware device using 8 bit word size) and the project team for which device and driver have been made did until now not complain.

Regards
Harald Mommer

On 05.03.24 08:46, Haixu Cui wrote:
Hello Harald,

Thank you for your detailed expatiation. To my knowledge, you took Vanilla as the front-end, and VMM is QEMU. Can you please explain further how do you test the SPI transfer without the Vanilla userspace interface? Thanks again.

Haixu Cui