[keyonjie]

The DMA channels need to be shared among pipelines and DSP cores,
allocate buffer for it in RUNTIME_SHARED area.

Signed-off-by: Keyon Jie yang.jie@linux.intel.com

[lyakh]

I'm wondering: do we know exactly how different cores access the same HDA DMA channel? One possibility is, that if the channel is allocated using rzalloc() on core 0 as a part of initialisation, then core 1 uses it exclusively. It initialises its various fields but then the original memset(0) from rzalloc() on core 0 writes caches back and overwrites core 1 data. If that is the case, maybe just adding one writeback after rzalloc() would be an easier fix?

[keyonjie]

I'm wondering: do we know exactly how different cores access the same HDA DMA channel? One possibility is, that if the channel is allocated using rzalloc() on core 0 as a part of initialisation, then core 1 uses it exclusively. It initialises its various fields but then the original memset(0) from rzalloc() on core 0 writes caches back and overwrites core 1 data. If that is the case, maybe just adding one writeback after rzalloc() would be an easier fix?

These channels is allocated in the first usage of a channel (on core 0, or a secondary core), and different channels can be used for pipelines run on different Cores. These hda_chan structs can be changed at runtime (e.g. change get/put), so allocate in in uncached area is much easier for us.

[lyakh]

I'm wondering: do we know exactly how different cores access the same HDA DMA channel? One possibility is, that if the channel is allocated using rzalloc() on core 0 as a part of initialisation, then core 1 uses it exclusively. It initialises its various fields but then the original memset(0) from rzalloc() on core 0 writes caches back and overwrites core 1 data. If that is the case, maybe just adding one writeback after rzalloc() would be an easier fix?

These channels is allocated in the first usage of a channel (on core 0, or a secondary core), and different channels can be used for pipelines run on different Cores. These hda_chan structs can be changed at runtime (e.g. change get/put), so allocate in in uncached area is much easier for us.

I think they're allocated in hda_dma_probe() once for all channels and that probably runs on core 0 at init time. Then I suspent that channels are allocated for specific purposes: for audio pipelines, for DMA trace etc. DMA trace is done from a thread, running on core 0. Audio pipelines also run on specific cores. So I actually would expect DMA channels at run time to be only used by the same core? PCMs and DAIs also belong to specific pipelines running on their respective cores. AFAICS DMA channels are also assigned in DAI and host configuration. So, sorry, I'm still not sure how channels can be used by different cores? Maybe we can add a comment somewhere to document this.

[lgirdwood]

CI showing known and unrelated issue.

[lgirdwood]

@lyakh I suspect the chan lookup/free probably needs the uncache alias as we may try and use a channel twice. TBC though.