Fix file chosing for CMA

[xR3b0rn]

This issue can be seen as a follow-up to libiio issue #1259.

According to this kernel issue, linux,cma is not a universally defined name - that is, it's not guaranteed to point to the CMA memory on every system. The actual name of the device file depends on the name of the CMA node in the device tree or may be reserved if cma=<size> is set on the linux cmdline.

So, technically speaking, the hardcoded path /dev/dma_heap/linux,cma may be correct on many systems, but it is not guaranteed to work everywhere. If cma=<size> is specified or a differently named device tree node is used, the corresponding file name will change. As such, the hardcoded path referenced here may not actually refer to the CMA on all systems.

[schwesinger]

FYI:

During the rebase of one of our drivers onto v6.17 I ran into new kernel config option DMABUF_HEAPS_CMA_LEGACY that might be relevant for deciding how to proceed with this in libiio:

[PATCH v3 0/3] dma-buf: heaps: Use constant name for CMA heap: https://lkml.org/lkml/2025/5/22/1368

Also there's this WIP patchset:

[PATCH v7 0/5] dma-buf: heaps: Create a CMA heap for each CMA reserved region: https://lkml.org/lkml/2025/7/21/626

I didn't have the time to dig any deeper, especially into the linked discussions in the ML threads.

refs #1344, #1339

[nunojsa]

From a quick look, I do not think it's too relevant for libiio. There's still (AFAICT) no way, at runtime, to inspect if a given device needs to allocate a CMA region or if it can use normal system heaps. So I think that having the env var is the less disruptive option for now.

Alternatively, I think it would be fairly safe to just try first to get CMA backed buffers and if that fails, fallback to system heaps. And we could still have the env variable to opt out from that behavior (in case we have something else needing the CMA memory and we do not want the IIO streaming device "stealing" that memory nor we can/want to make the region extremely big to accommodate both)

[xR3b0rn]

Thanks for still working on this!

Just my humble impression: I feel like CMA might be a fairly limited resource on some systems, so using it as the default doesn’t seem ideal.

[nunojsa]

Just my humble impression: I feel like CMA might be a fairly limited resource on some systems, so using it as the default doesn’t seem ideal.

I know, and it's a very uncommon case. That's why I would default to use it when available (if it's there, probably there's a very strong reason for it). We would not fail if CMA is not there, we would just default for the more common case. I know this would be one more syscall but it's also not a fastpath at all so it would be more than acceptable. But this is just me throwing things to the air. At this point, I guess the most "safe" thing to do is the env variable and do not try to guess "usecases".