Re: [PATCH 5/5] cxl: Avoid to create dax regions for type2 accelerators

[Dan Williams]

Huang Ying wrote:
> The memory range of a type2 accelerator should be managed by the type2
> accelerator specific driver instead of the common dax region drivers,
> as discussed in [1].
> 
> [1] https://lore.kernel.org/linux-cxl/66469ff1b8fbc_2c2629427@xxxxxxxxxxxxxxxxxxxxxxxxx.notmuch/
> 
> So, the patch skips dax regions creation for type2 accelerator device
> memory regions.
> 
> Based on: https://lore.kernel.org/linux-cxl/168592159835.1948938.1647215579839222774.stgit@xxxxxxxxxxxxxxxxxxxxxxxxx/
> 
> Signed-off-by: "Huang, Ying" <ying.huang@xxxxxxxxx>
> Co-developed-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> Reviewed-by: Gregory Price <gourry@xxxxxxxxxx>
> Cc: Davidlohr Bueso <dave@xxxxxxxxxxxx>
> Cc: Jonathan Cameron <jonathan.cameron@xxxxxxxxxx>
> Cc: Dave Jiang <dave.jiang@xxxxxxxxx>
> Cc: Alison Schofield <alison.schofield@xxxxxxxxx>
> Cc: Vishal Verma <vishal.l.verma@xxxxxxxxx>
> Cc: Ira Weiny <ira.weiny@xxxxxxxxx>
> Cc: Alejandro Lucero <alucerop@xxxxxxx>
> Cc: Ben Cheatham <benjamin.cheatham@xxxxxxx>
> ---
>  drivers/cxl/core/region.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c
> index d709738ada61..708be236c9a2 100644
> --- a/drivers/cxl/core/region.c
> +++ b/drivers/cxl/core/region.c
> @@ -3473,6 +3473,14 @@ static int cxl_region_probe(struct device *dev)
>  					p->res->start, p->res->end, cxlr,
>  					is_system_ram) > 0)
>  			return 0;
> +		/*
> +		 * Accelerator regions have specific usage, skip
> +		 * device-dax registration.
> +		 */
> +		if (cxlr->type == CXL_DECODER_ACCEL)
> +			return 0;
> +
> +		/* Expander routes to device-dax */

So Linux is this weird career choice where you get to argue with
yourself months later. I understand why I took this shortcut of assuming
that the coherence mode determines device-dax routing, but that is not
sufficient.

An HDM-DB region could want the device-dax uAPI, and an HDM-H range
could want to do its own uAPI besides device-dax.

So, in retrospect, I think it is a mistake to assume uAPI from coherence
mode. It really is a property of the region decided by the region
creator independent of the coherence mode or the device type.

I am thinking that 'struct cxl_region' grows something like:

diff --git a/drivers/cxl/cxl.h b/drivers/cxl/cxl.h
index 5406e3ab3d4a..4cf1d030404d 100644
--- a/drivers/cxl/cxl.h
+++ b/drivers/cxl/cxl.h
@@ -511,12 +511,25 @@ struct cxl_region_params {
  */
 #define CXL_REGION_F_NEEDS_RESET 1
 
+/*
+ * enum cxl_mem_api - where to route this cxl region
+ * @CXL_MEM_API_DAX: application specific / soft-reserved memory
+ * @CXL_MEM_API_PMEM: direct region to the NVDIMM subsystem
+ * @CXL_MEM_API_NONE: accelerator specific / hard-reserved memory
+ */
+enum cxl_mem_api {
+	CXL_MEM_API_DAX,
+	CXL_MEM_API_PMEM,
+	CXL_MEM_API_NONE,
+};
+
 /**
  * struct cxl_region - CXL region
  * @dev: This region's device
  * @id: This region's id. Id is globally unique across all regions
  * @mode: Endpoint decoder allocation / access mode
  * @type: Endpoint decoder target type
+ * @api: What if any subsystem will present this region to consumers
  * @cxl_nvb: nvdimm bridge for coordinating @cxlr_pmem setup / shutdown
  * @cxlr_pmem: (for pmem regions) cached copy of the nvdimm bridge
  * @flags: Region state flags
@@ -530,6 +543,7 @@ struct cxl_region {
 	int id;
 	enum cxl_decoder_mode mode;
 	enum cxl_decoder_type type;
+	enum cxl_mem_api api;
 	struct cxl_nvdimm_bridge *cxl_nvb;
 	struct cxl_pmem_region *cxlr_pmem;
 	unsigned long flags;

Now, I have not seen how Alejandro's series handles this, but this
type-2 series was shorter so I started here first.