On Thu, Jan 20, 2022 at 03:03:40PM +0100, Christoph Hellwig wrote:

On Wed, Jan 12, 2022 at 06:37:03PM +0000, Matthew Wilcox wrote:

...

But let's go further than that (which only brings us to 32 bytes per range). For the systems you care about which use an identity mapping, and have sizeof(dma_addr_t) == sizeof(phys_addr_t), we can simply point the dma_range pointer to the same memory as the phyr. We just have to not free it too early. That gets us down to 16 bytes per range, a saving of 33%.

Even without an IOMMU the dma_addr_t can have offsets vs the actual physical address. Not on x86 except for a weirdo SOC, but just about everywhere else.

The point is dma_map knows if that is happening or not and giving dma_map the option to just return a pointer to the input memory to re-use as the dma list does optimize important widely used cases.

Yes, some weirdo SOC cannot do this optimization, but the weirdo SOC will allocate a new memory and return the adjusted dma_addr_t just fine.

Ideally we should not pay a cost for weirdo SOC on sane systems.

Jason