zephyr: make DMA LLI allocations always uncached

[lyakh]

Extensive testing showed, that making memory allocations from the SOF_MEM_ZONE_RUNTIME zone uncached fixes bug #4414 on APL with Zephyr. We plan to introduce a clean separation between cached and uncached memory access modes, but this is a significant change, planned for the next SOF release. Until then we disable cached runtime allocations in Zephyr.

[paulstelian97]

@dbaluta Does this affect our pipeline performance, especially for decoders (which might have state in the RUNTIME zone)? Just a FYI, I guess we can allow this to go and locally revert if it is a problem...

[lyakh]

A dedicated manual CI run with this PR triggered the bug again. Making this a draft for further investigation.

[marcinszkudlinski]

anyway, isn't too drastic sollution? Turning off the cache may

1. (obvious) make significant performance drop
2. (much worse and much more dangerous) lead to many cache related defect stay undetected
   Stabilizing the firmware after turning the cache back on may take weeks

[lyakh]

anyway, isn't too drastic sollution? Turning off the cache may

1. (obvious) make significant performance drop

2. (much worse and much more dangerous) lead to many cache related defect stay undetected
   Stabilizing the firmware after turning the cache back on may take weeks

@marcinszkudlinski (1) performance - we will compare once we have a working solution. Zephyr reports load statistics, so should be easy to check. (2) regressions - as the commit message says, we will enforce correct cache usage by adding compile-time checks, using sparse and maybe the compiler itself. That should really clean up most errors in that area

[marcinszkudlinski]

I would't say "bad fix" as the title says. Cached or uncached - in fact it is the very same heap and aligment always should be kept to the cacheline.

[lyakh]

SOFCI TEST

[lgirdwood]

SOFCI TEST

[lgirdwood]

Re test with the single commit.

[lyakh]

CI: three Zephyr platforms failed individual cases with "sof-logger dead" and a single non-Zephyr TGLU platform failed 2 suspend-resume tests with audio