On Fri, Apr 23, 2021 at 05:20:55PM +0100, Mark Brown wrote: > Part of the issue here is that there has been some variation in how > num_chipselect is interpreted with regard to GPIO based chip selects > over time. It *should* be redundant, I'm not clear why it's in the > generic bindings at all but that's lost to history AFAICT. It seems num_chipselect is meant to be set to the maximum number of *native* chipselects supported by the controller. Which is overwritten if GPIO chipselects are used. I failed to appreciate that when I changed num_chipselects for spi-bcm2835.c with commit 571e31fa60b3. That single line change in the commit ought to be reverted. And the kernel-doc ought to be amended because the crucial detail that num_chipselect needs to be set to the maximum *native* chipselects isn't mentioned anywhere. > The best thing would be to have it not have a single array of chip > select specific data and instead store everything in the controller_data > that's there per-device. Unfortunately that's non-trivial. The slave-specific data is DMA-mapped. It could be DMA-mapped in ->setup but there's no ->unsetup to DMA-unmap the memory once the slave is removed. Note that the slave could be removed dynamically with a DT overlay, not just when the controller is unbound. So we'd need a new ->unsetup hook at the very least to make this work. The Foundation's downstream kernel now contains a bandaid commit which raises the limit to 24 and errors out of ->probe if the limit is exceeded. I would have preferred if it errored out of ->setup. That way only the slaves exceeding the limit would fail to instantiate: https://github.com/raspberrypi/linux/commit/05f8d5826e28 Thoughts? Lukas