[PATCH v3 2/2] block-dma: properly take MMIO path

[Leon Romanovsky]

On Mon, Oct 27, 2025 at 08:49:22AM +0100, Christoph Hellwig wrote:
> > +	switch (iter.p2pdma.map) {
> > +	case PCI_P2PDMA_MAP_BUS_ADDR:
> > +		iod->flags |= IOD_DATA_P2P;
> > +		break;
> > +	case PCI_P2PDMA_MAP_THRU_HOST_BRIDGE:
> > +		iod->flags |= IOD_DATA_MMIO;
> > +		break;
> > +	default:
> > +		return BLK_STS_RESOURCE;
> 
> I almost wonder if we should just the pci_p2pdma_map_type values into
> place.  But that's a future cleanup, I'd rather get this going now.

I thought about it, but decided to use flags as more space efficient.
PCI_P2PDMA_MAP_BUS_ADDR == 3 and PCI_P2PDMA_MAP_THRU_HOST_BRIDGE == 4.
It means that will need to occupy 6 bits (extra u8 IOD field), while
this flags encoding takes only 4 bits (without extra IOD field).

> 
> > +static inline bool blk_rq_dma_unmap(struct request *req, struct device *dma_dev,
> > +		struct dma_iova_state *state, size_t mapped_len,
> > +		enum pci_p2pdma_map_type map)
> >  {
> > -	if (is_p2p)
> > +	if (map == PCI_P2PDMA_MAP_BUS_ADDR)
> >  		return true;
> >  
> >  	if (dma_use_iova(state)) {
> > +		unsigned int attrs = 0;
> > +
> > +		if (map == PCI_P2PDMA_MAP_THRU_HOST_BRIDGE)
> > +			attrs |= DMA_ATTR_MMIO;
> > +
> >  		dma_iova_destroy(dma_dev, state, mapped_len, rq_dma_dir(req),
> > -				 0);
> > +				 attrs);
> 
> The only thing in req that is used now is the data directrion.  I'd be
> almost tempted to just pass that and lift this to dma-mapping.h.
> 
> But I guess we could just do that in a follow on to not drag in
> another subsystem.
> 
> Otherwise looks good:
> 
> Reviewed-by: Christoph Hellwig <hch at lst.de>

Thanks