Re: [PATCH v4 14/16] block-dma: migrate to dma_map_phys instead of map_page

[Jason Gunthorpe <jgg@xxxxxxxxxx>]

On Tue, Sep 02, 2025 at 03:59:37PM -0600, Keith Busch wrote:
> On Tue, Sep 02, 2025 at 10:49:48PM +0200, Marek Szyprowski wrote:
> > On 19.08.2025 19:36, Leon Romanovsky wrote:
> > > @@ -87,8 +87,8 @@ static bool blk_dma_map_bus(struct blk_dma_iter *iter, 
> > > struct phys_vec *vec)
> > >   static bool blk_dma_map_direct(struct request *req, struct device 
> > > *dma_dev,
> > >                   struct blk_dma_iter *iter, struct phys_vec *vec)
> > >   {
> > > - iter->addr = dma_map_page(dma_dev, phys_to_page(vec->paddr),
> > > -                 offset_in_page(vec->paddr), vec->len, rq_dma_dir(req));
> > > + iter->addr = dma_map_phys(dma_dev, vec->paddr, vec->len,
> > > +                 rq_dma_dir(req), 0);
> > >           if (dma_mapping_error(dma_dev, iter->addr)) {
> > >                   iter->status = BLK_STS_RESOURCE;
> > >                   return false;
> > 
> > I wonder where is the corresponding dma_unmap_page() call and its change 
> > to dma_unmap_phys()...
> 
> You can't do that in the generic layer, so it's up to the caller. The
> dma addrs that blk_dma_iter yield are used in a caller specific
> structure. For example, for NVMe, it goes into an NVMe PRP. The generic
> layer doesn't know what that is, so the driver has to provide the
> unmapping.

To be specific I think it is this hunk in another patch that matches
the above:

@@ -682,11 +682,15 @@ static void nvme_free_prps(struct request *req)
 {
        struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
        struct nvme_queue *nvmeq = req->mq_hctx->driver_data;
+       unsigned int attrs = 0;
        unsigned int i;
 
+       if (req->cmd_flags & REQ_MMIO)
+               attrs = DMA_ATTR_MMIO;
+
        for (i = 0; i < iod->nr_dma_vecs; i++)
-               dma_unmap_page(nvmeq->dev->dev, iod->dma_vecs[i].addr,
-                               iod->dma_vecs[i].len, rq_dma_dir(req));
+               dma_unmap_phys(nvmeq->dev->dev, iod->dma_vecs[i].addr,
+                               iod->dma_vecs[i].len, rq_dma_dir(req), attrs);


And it is functionally fine to split the series like this because
unmap_page is a nop around unmap_phys:

void dma_unmap_page_attrs(struct device *dev, dma_addr_t addr, size_t size,
                 enum dma_data_direction dir, unsigned long attrs)
{
        if (unlikely(attrs & DMA_ATTR_MMIO))
                return;

        dma_unmap_phys(dev, addr, size, dir, attrs);
}
EXPORT_SYMBOL(dma_unmap_page_attrs);

Jason