DMA mapping fill dma_address to 0

[Corentin Labbe <clabbe.montjoie@gmail.com>]

Hello

I work on the crypto offloader driver of cortina/gemini SL3516 SoC.
I test it by filling a LUKS2 partition.
I got a reproductible problem when handling skcipher requests.
I use dma_map_sg() and when iterating other the result, sg_dma_address(sg) return 0.
But sg_dma_len(sg) is still correct (4096 in my case).

Below is a simplified view of my code:
nr_sgs = dma_map_sg(ce->dev, areq->src, sg_nents(areq->src), DMA_TO_DEVICE);
(nr_sgs = 1 in my case)
sg = areq->src;
if (!sg_dma_address(sg))
	FAIL

I have digged to find what do dma_map_sg() and I have added some debug.
sg_page(sg) return c7efb000 for example so sg_page() works.
But it seems the problem is that page_to_phys(sg_page(sg)) return 0.

This problem does not appear immediatly, luksOpen and subsequent fsck always work.
But it appears fast after, when mouting or rsync files in it.

I have added CONFIG_DEBUG_SG, CONFIG_DMA_API_DEBUG, CONFIG_DMA_API_DEBUG_SG but they didnt bringed any more hints.
Only "DMA-API: cacheline tracking ENOMEM, dma-debug disabled" appears but always with some "time" between my problem and its display.
So I am not sure it is related.

Regards