Re: [PATCH v9 04/10] cxl: Add HDM-DB region creation

[Alison Schofield <alison.schofield@intel.com>]

On Tue, Sep 22, 2026 at 04:38:41PM -0700, Davidlohr Bueso wrote:
> A region inherits its coherency from the chosen root decoder - HDM-DB
> if the root has CXL_DECODER_F_BI, otherwise HDM-H.

Above holds for regions created from sysfs. An auto region takes its type
from the first committed endpoint decoder in construct_region(), without
looking at the root decoder. A committed host-only decoder in a BI window
becomes an HDM-H region under a BI root, and an HDM-D region is neither
HDM-DB nor HDM-H. Can this say which regions it applies to?

> 
> cxl_acpi_cfmws_verify() rejects a Window that declares no coherency
> model at all (neither Device Coherent nor Host-only Coherent), one that
> sets BI together with Host-only Coherent, which the CFMWS definition
> calls undefined behavior, and one that sets BI without Device Coherent,
> since HDM-DB is defined only as bit[0] and bit[5] together. A BI Window
> therefore always exposes device-coherent memory and nothing else.

These rejections drop windows that are accepted and usable today.  See
the cxl_acpi_cfmws_verify() comment inline below.

> The root decoder's target_type follows the Window as well -
> device-coherent when only Device Coherent is set, host-only otherwise.

I'll comment inline on this one too, acpi.c hunk below. I think this
describes dead code and opportunity for cleanup.

> 
> Introduce the following read-only sysfs ABI.
> 
>   - decoderX.Y/cap_back_invalidate (root) reports the CFMWS BI
>     restriction.
>   - decoderX.Y/back_invalidate (endpoint) reads '1' when configured
>     for HDM-DB.
> 
> cxl_region_attach() rejects endpoints whose device or HDM cannot
> serve the region's type; target_type is inherited from cxlr->type
> in cxl_rr_assign_decoder(), restored to the endpoint default on
> detach, and to whatever it was on a failed attach - a refusal may
> come before any inheritance, for a decoder another region owns or
> one firmware committed, and must not relabel it.

A decoder firmware committed is still relabelled when its attach succeeds,
since the type mismatch check is gone. More inline at
cxl_region_attach()

snip
> +
> +
> +What:		/sys/bus/cxl/devices/decoderX.Y/back_invalidate
> +Date:		September, 2026
> +KernelVersion:	v7.4
> +Contact:	linux-cxl@vger.kernel.org
> +Description:
> +		(RO) Shows '1' if this endpoint decoder is currently configured
> +		for HDM-DB (device-managed coherency with back-invalidate).
> +		The HDM-DB state is inherited from the region the decoder is
> +		attached to, which is in turn set from the chosen root
> +		decoder's CFMWS BI restriction (see cap_back_invalidate).
> +
>  What:		/sys/bus/cxl/devices/decoderX.Y/delete_region
>  Date:		May, 2022
>  KernelVersion:	v6.0
> diff --git a/drivers/cxl/acpi.c b/drivers/cxl/acpi.c
> index 3b818adbd38b..2e8e31544a5f 100644
> --- a/drivers/cxl/acpi.c
> +++ b/drivers/cxl/acpi.c
> @@ -152,6 +152,8 @@ static unsigned long cfmws_to_decoder_flags(int restrictions)
>  		flags |= CXL_DECODER_F_PMEM;
>  	if (restrictions & ACPI_CEDT_CFMWS_RESTRICT_FIXED)
>  		flags |= CXL_DECODER_F_LOCK;
> +	if (restrictions & ACPI_CEDT_CFMWS_RESTRICT_BI)
> +		flags |= CXL_DECODER_F_BI;
>  
>  	return flags;
>  }
> @@ -198,6 +200,24 @@ static int cxl_acpi_cfmws_verify(struct device *dev,
>  		dev_dbg(dev, "CFMWS length %d greater than expected %d\n",
>  			cfmws->header.length, expected_len);
>  
> +	if ((cfmws->restrictions & ACPI_CEDT_CFMWS_RESTRICT_HOSTONLYMEM) &&
> +	    (cfmws->restrictions & ACPI_CEDT_CFMWS_RESTRICT_BI)) {
> +		dev_err(dev, "CFMWS cannot have both HDM-H and HDM-DB\n");
> +		return -EINVAL;
> +	}
> +
> +	if (!(cfmws->restrictions & (ACPI_CEDT_CFMWS_RESTRICT_DEVMEM |
> +				     ACPI_CEDT_CFMWS_RESTRICT_HOSTONLYMEM))) {
> +		dev_err(dev, "CFMWS has no coherency model\n");
> +		return -EINVAL;
> +	}
> +
> +	if ((cfmws->restrictions & ACPI_CEDT_CFMWS_RESTRICT_BI) &&
> +	    !(cfmws->restrictions & ACPI_CEDT_CFMWS_RESTRICT_DEVMEM)) {
> +		dev_err(dev, "CFMWS BI requires device-coherent\n");
> +		return -EINVAL;
> +	}
> +


Can these checks break windows that are accepted and usable today?  A
rejected window gets no root decoder, and auto-assembly then fails with
"no CXL window for range".

A window that sets neither coherency bit is accepted today and can hold
auto regions, since assembly never looks at the restriction flags.

A window that sets BI together with Host-only, or BI without Device Coherent,
is accepted today as a non-BI window, because BI is ignored.
Manual HDM-H region creation in it is lost as well.

Even if no shipping firmware does this, is cxl_acpi_cfmws_verify() the right
place for these checks? Til now it has only rejected windows that cannot be
decoded (arithmetic, alignment, ways, length).  What a window may be used for
is decided later, by cfmws_to_decoder_flags() and can_create_ram|pmem(). 
Might it be enough to withhold CXL_DECODER_F_BI for an invalid combination,
so that only HDM-DB is refused?


>  	return 0;
>  }
>  
> @@ -437,7 +457,14 @@ static int __cxl_parse_cfmws(struct acpi_cedt_cfmws *cfmws,
>  
>  	cxld = &cxlrd->cxlsd.cxld;
>  	cxld->flags = cfmws_to_decoder_flags(cfmws->restrictions);
> +	/* host-only wins if firmware sets both coherency restrictions */
>  	cxld->target_type = CXL_DECODER_HOSTONLYMEM;
> +	if (cxld->flags & CXL_DECODER_F_TYPE2) {
> +		if (cxld->flags & CXL_DECODER_F_TYPE3)
> +			dev_dbg(dev, "CFMWS has both HDM-H and HDM-D\n");
> +		else
> +			cxld->target_type = CXL_DECODER_DEVMEM;
> +	}


Mentioned in commit msg, I think this is dead code needing cleanup.
I don't see a consumer of a root decoder's target_type.

Building window-derived logic and a dev_dbg() on top of it makes it
look as if the root decoder's type matters when it doesn't.  Can this
hunk be dropped?  Removing the old line can be a separate cleanup.


>  	cxld->hpa_range = (struct range) {
>  		.start = cfmws->base_hpa,
>  		.end = cfmws->base_hpa + cfmws->window_size - 1,
> diff --git a/drivers/cxl/core/hdm.c b/drivers/cxl/core/hdm.c
> index 18200a8f3f72..44ff64b1a9df 100644
> --- a/drivers/cxl/core/hdm.c
> +++ b/drivers/cxl/core/hdm.c
> @@ -705,9 +705,21 @@ static void cxld_set_interleave(struct cxl_decoder *cxld, u32 *ctrl)
>  
>  static void cxld_set_type(struct cxl_decoder *cxld, u32 *ctrl)
>  {
> +	bool bi = cxld->target_type == CXL_DECODER_DEVMEM &&
> +		  cxld->region && cxl_root_decoder_is_bi(cxld->region->cxlrd);
> +
>  	u32p_replace_bits(ctrl,
>  			  !!(cxld->target_type == CXL_DECODER_HOSTONLYMEM),
>  			  CXL_HDM_DECODER0_CTRL_HOSTONLY);
> +	u32p_replace_bits(ctrl, bi, CXL_HDM_DECODER0_CTRL_BI);
> +
> +	if (bi && is_endpoint_decoder(&cxld->dev)) {
> +		struct cxl_endpoint_decoder *cxled =
> +			to_cxl_endpoint_decoder(&cxld->dev);
> +
> +		u32p_replace_bits(ctrl, cxled->pos,
> +				  CXL_HDM_DECODER0_CTRL_ISP_MASK);
> +	}


The commit message does not mention ISP programming. Can it get a
sentence with the spec reference for ISP being required when BI is set?

snip

> diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c

snip

> @@ -1131,16 +1163,11 @@ static int cxl_rr_assign_decoder(struct cxl_port *port, struct cxl_region *cxlr,
>  	}
>  
>  	/*
> -	 * Endpoints should already match the region type, but backstop that
> -	 * assumption with an assertion. Switch-decoders change mapping-type
> -	 * based on what is mapped when they are assigned to a region.
> +	 * Endpoint decoders inherit their type from cxlr->type; broken
> +	 * pairings were already rejected by the coherency checks in
> +	 * cxl_region_attach(). Switch-decoders change mapping-type based
> +	 * on what is mapped when they are assigned to a region.
>  	 */
> -	dev_WARN_ONCE(&cxlr->dev,
> -		      port == cxled_to_port(cxled) &&
> -			      cxld->target_type != cxlr->type,
> -		      "%s:%s mismatch decoder type %d -> %d\n",
> -		      dev_name(&cxled_to_memdev(cxled)->dev),
> -		      dev_name(&cxld->dev), cxld->target_type, cxlr->type);
>  	cxld->target_type = cxlr->type;


A Type 3 endpoint decoder in an HDM-DB region now reads "accelerator" from
decoderX.Y/target_type, and the ndctl test on the branch linked in the cover
letter checks for that. So I take it target_type is now being used to report
the decoder's coherency model, device-coherent vs host-only, rather than the
Type-2/Type-3 as documented by the existing ABI.

Does this need sysfs ABI update?

cxl list doesn't show target_type, but libcxl exports it through
cxl_decoder_get_target_type().  If that is intentional, libcxl.txt will
need the same semantic update with the ndctl patches.


>  	cxl_rr->decoder = cxld;
>  	return 0;
> @@ -1803,6 +1830,7 @@ static int cxl_region_attach_position(struct cxl_region *cxlr,
>  	struct cxl_root_decoder *cxlrd = cxlr->cxlrd;
>  	struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
>  	struct cxl_switch_decoder *cxlsd = &cxlrd->cxlsd;
> +	enum cxl_decoder_type type = cxled->cxld.target_type;
>  	struct cxl_decoder *cxld = &cxlsd->cxld;
>  	int iw = cxld->interleave_ways;
>  	struct cxl_port *iter;
> @@ -1828,6 +1856,8 @@ static int cxl_region_attach_position(struct cxl_region *cxlr,
>  	for (iter = cxled_to_port(cxled); !is_cxl_root(iter);
>  	     iter = to_cxl_port(iter->dev.parent))
>  		cxl_port_detach_region(iter, cxlr, cxled);
> +	/* undo cxl_rr_assign_decoder() type inheritance */
> +	cxled->cxld.target_type = type;
>  	return rc;
>  }
>  
> @@ -2056,6 +2086,7 @@ static int cxl_region_attach(struct cxl_region *cxlr,
>  	struct cxl_region_params *p = &cxlr->params;
>  	struct cxl_port *ep_port, *root_port;
>  	struct cxl_dport *dport;
> +	struct cxl_hdm *cxlhdm;
>  	int rc = -ENXIO;
>  
>  	rc = check_interleave_cap(&cxled->cxld, p->interleave_ways,
> @@ -2105,10 +2136,39 @@ static int cxl_region_attach(struct cxl_region *cxlr,
>  		return -ENXIO;
>  	}
>  
> -	if (cxled->cxld.target_type != cxlr->type) {
> -		dev_dbg(&cxlr->dev, "%s:%s type mismatch: %d vs %d\n",
> -			dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
> -			cxled->cxld.target_type, cxlr->type);


Should the type check stay for committed decoders?  With it gone, a
second auto-assembled decoder whose committed Target Range Type differs
from the region's is accepted, and cxl_rr_assign_decoder() then
overwrites its target_type while the hardware keeps the committed value.
I think this changes in Patch 8 - which really makes patch 8 required,
not optional. (more on that in patch 8)

Maybe the CXL_DECODER_STATE_AUTO mismatch check fits better in here,
since this patch that removes the old one?


> +	/*
> +	 * Verify the device and HDM are capable of the region's flavor before
> +	 * proceeding. The endpoint decoder's target_type is then inherited
> +	 * from cxlr->type later in cxl_rr_assign_decoder().
> +	 */
> +	if (cxlr->type == CXL_DECODER_DEVMEM &&
> +	    cxl_root_decoder_is_bi(cxlrd) && !cxlds->bi) {
> +		dev_err(&cxlr->dev, "%s:%s BI not enabled on device\n",
> +			dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev));
> +		return -ENXIO;
> +	}
> +
> +	if (cxled->state != CXL_DECODER_STATE_AUTO &&
> +	    cxlr->type == CXL_DECODER_HOSTONLYMEM &&
> +	    cxlds->type == CXL_DEVTYPE_DEVMEM) {
> +		dev_warn(&cxlr->dev, "%s:%s HDM-H requires a Type 3 device\n",
> +			 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev));
> +		return -ENXIO;
> +	}
> +
> +	cxlhdm = dev_get_drvdata(&ep_port->dev);
> +	if (!cxlhdm)
> +		return -ENXIO;
> +	if (cxlr->type == CXL_DECODER_HOSTONLYMEM &&
> +	    cxlhdm->supported_coherency == CXL_HDM_DECODER_COHERENCY_DEV) {
> +		dev_warn(&cxlr->dev, "%s:%s HDM is device-coherent only\n",
> +			 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev));
> +		return -ENXIO;
> +	}
> +	if (cxlr->type == CXL_DECODER_DEVMEM &&
> +	    cxlhdm->supported_coherency == CXL_HDM_DECODER_COHERENCY_HOST) {
> +		dev_warn(&cxlr->dev, "%s:%s HDM is host-only coherent\n",
> +			 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev));
>  		return -ENXIO;
>  	}


Why dev_err()/dev_warn() here?  The type mismatch refusal this replaces was
a dev_dbg().  The attach fail here would be reported as -ENXIO.

Tidy up opportunity, maybe move the 4 coherency checks into a helper so
cxl_region_attach() stays cleaner.

Another cleanup - the 'is this reion HDM-DB test is now coded in a few
places, my look may have been limited:
	cxld_set_type():		target_type == DEVMEM && cxl_root_decoder_is_bi()
	back_invalidate_show():		same, plus cxlds->bi
	cxl_region_decode_commit():	cxlr->type == DEVMEM && cxl_root_decoder_is_bi()
	cxl_region_attach():		same

Maybe a single cxl_region_is_hdm_db() would be clearer and cleaner.



snip to end