The open() method is not well suited to hot-plugged SPI device controllers (e.g. USB, Thunderbolt, PCIe and similar situations).
tim-seoss opened this issue · comments
The Linux kernel usually doesn't guarantee that SPI bus numbers are predictable and persistent across reboots, and kernel upgrades etc.
As with other device special files (e.g. storage devices, network devices, serial ports), the Linux udev subsystem is the generic mechanism for creating persistent/predictable names for devices (including SPI bus numbers).
Whilst in many cases the SPI bus number is predicable (e.g. Raspberry Pi OS maintainers take steps to ensure the spi bus number remains stable), in general this is not the case (e.g. Beaglebone Black has swapped SPI bus numbers 0 and 1 more than once between major Linux kernel version upgrades).
Unfortunately the current open() method requires that the user of the extension knows the bus number because it hard-codes the use of /dev/spidevX.Y.
A solution would be to add a second alternative call e.g. open_path() which would allow the user to supply a persistent device name such as /dev/spi/by-path/ftdi232h_7821067-cs0. Such a filesystem path would typically be created automatically by the udev daemon when, and would take the form of a symbolic link to a (non-predictable) spidev character device file e.g. /dev/spidev/14.0.
Having studied the codebase, it would be possible to keep the existing open() call (for backward compatibility), whilst adding a more generic open_path() alternative.
I'll prepare a PR with a draft implementation, along with some sample udev rules, and documentation updates.
This is a good point, and thank you for your PR.
I need to find the time to test this, since I am able to accept PRs on my fork and release a new version of the spidev Python library if it makes sense.
I'm currently fighting bugs on Pi 5, however, but after that I'll be updating some libraries that use spidev and it would make sense for me to test changes like this then.
Thanks, any feedback is welcome. An organisation I work with has put this code into production, so it's received a reasonably amount of testing on a couple of different 32 bit ARM systems. I've also used it with USB to SPI adapters on x86-64.