Host-managed device memory decoder
An HDM decoder describes address ranges and interleaving used to route accesses to CXL device memory or translate them at an endpoint.
In the Linux 6.17 decoder model, root and switch decoders route host physical addresses (HPA); endpoints translate them into device physical addresses (DPA). Root decoders represent platform memory windows and need not be separate hardware blocks. [1]
Interleave ways and granularity distribute addresses among targets. Root and switch settings describe their immediate downstream targets; endpoint and region settings describe the entire interleave set. These settings must be consistent, but need not be identical at every level. [1]
Linux 6.17 can expose /dev/daxN.Y for application-managed mmap() access or convert capacity to page-allocator memory through dax_kmem. [2] The underlying memory may still be volatile. [3]
Relations
- related
- CXL memory expansion, CXL Type-3 device, Direct Access, Byte-addressable storage, CXL memory pooling
Citations
- Linux 6.17: CXL Driver Operation, Decoders and Regions.
- Linux 6.17: DAX Driver Operation, device DAX and dax_kmem.
- Linux 6.17: Devices and Protocols, CXL memory and Type-3 devices.
Sources & caveats
- Decoder taxonomy follows the version-pinned Linux 6.17 driver model. Endpoint-only HPA-to-DPA translation is an explicit Linux assumption; this is not a universal platform support claim. [1]
- The lifecycle field is unknown because a decoder mechanism has no independently established product lifecycle.