Re: [RFC PATCH 1/4] cxl/extent: Promote cxlr_dax->region_extent to an xarray

[Gregory Price <gourry@gourry.net>]

On Wed, Apr 29, 2026 at 11:56:38AM +0100, Jonathan Cameron wrote:
> On Mon, 27 Apr 2026 10:12:02 -0500
> Ira Weiny <ira.weiny@intel.com> wrote:
> 
> > Jonathan Cameron wrote:
> > >   
> > > > Look at the diagrams in this presentation:
> > > > 
> > > > https://lpc.events/event/18/contributions/1826/attachments/1435/3335/LPC2024_CXL_DCD-v2.pdf
> > > > 
> > > > 'DAX dev 1' covers memory from Extent A and Extent B.  What yall will want
> > > > to do is ensure that the region extents which get surfaced are ordered
> > > > based on the sequence number _when_ _the_ _dax_ _device_ is created.  The
> > > > order they come into the host does not really matter.  Although yea the
> > > > spec has a bunch of rules...  so whatever, follow those.  But it is the
> > > > dax device which groups the extents into a contiguous HPA range and maps  
> > > 
> > > On this bit, HPA?  Why would extents be contiguous in HPA?  Contiguous
> > > in the DAX device mapping sure, but not HPA.  
> > 
> > I'm not saying the _extents_ are contiguous in HPA.  I said they get
> > _mapped_ into a contiguous HPA _by_ the DAX device.
> 
> That's not HPA at that point, it's a VA of some type.
> Now maybe it smells like HPA in some DAX interface but if it does
> we should think about any renames etc necessary to avoid it doing
> so!  Or introduce DAXPA (not DPA for obvious reasons ;)
>

Dynamic Capacity Extent
Byte Offset      Length      Description
00h              08h         Starting DPA

This is all we need to know what should happen in the rest of the path.

The device sends a set of DPA-extents that map into a host defined DCD
region (which has an HPA range).  

The extents it sends may be laid out as such:

Extent0: Tag(A) Seq(0) DPA(0x0000.0000)  Length(0x1000.0000)
Extent1: Tag(A) Seq(1) DPA(0xF000.0000)  Length(0x1000.0000)
(start of device, end of device)

The CXL driver creates a DCD Region, which maps a *possible region of
memory* for extents, but this region *may be sparse*.

For example:

DCDRegion -> Base(0x5.0000.0000)  Length(0x1.0000.0000)
Extent0   -> Base(0x5.0000.0000)  Length(0x1000.0000)  (start of region)
Extent1   -> Base(0x5.F000.0000)  Length(0x1000.0000)  (end of region)

This is a completely valid set of extents.

We use DAX to cobble these sparse extents into a "virtually contiguous"
region of memory - essentially base-0 + length.

DAX0.0: Base(0x0) Length(0x2000.0000)
   Region List:
        Base(0x0)         -> Extent0
	Base(0x1000.0000) -> Extent1


   So in a single diagram:


   [       device dc partition      ]
   [extent0] ... unalloc... [extent1]    Device View
       |         Sparse!        |
----------------------------------------
       |         Sparse!        |
   [extent0] ... unalloc... [extent1]
   [         host dc region         ]    CXL Driver View
       |                        |
----------------------------------------
       |                        |
   [            DAX0.0              ]
   [   Extent0     ][    Extent1    ]    DAX View
       |       NOT Sparse      |
----------------------------------------
   fd = open("/dev/dax0.0")
   buf = mmap(fd, ...)                  Software View

   buf[0];           -> DPA 0x0000.0000
   buf[0x10000000];  -> DPA 0xF000.0000

~Gregory