Re: WARN_ON(irqs_disabled()) in dma_free_attrs?

[Christoph Hellwig <hch-jcswGhMUV9g@public.gmane.org>]

On Tue, Mar 13, 2018 at 12:11:49PM +0000, Robin Murphy wrote:
> Taking a step back, though, provided the original rationale about 
> dma_declare_coherent_memory() is still valid, I wonder if we should simply 
> permit the USB code to call dma_{alloc,free}_from_dev_coherent() directly 
> here instead of being "good" and indirecting through the top-level DMA API 
> (which is the part which leads to problems). Given that it's a specific DMA 
> bounce buffer implementation within a core API, not just any old driver 
> code, I personally would consider that reasonable.

Looking back I don't really understand why we even indirect the "classic"
per-device dma_declare_coherent_memory use case through the DMA API.

It seems like a pretty different use case to me.  In the USB case we
also have the following additional twist in that it doesn't even need
the mapping to be coherent.

So maybe for now the quick fix is to move the sleep check as suggested
earlier in this thread, but in the long run we probably need to do some
major rework of how dma_declare_coherent_memory and friends work.