Message-ID: <11544820-6228-4930-8849-2121c6b0a3b7@intel.com>
Date: Mon, 6 Oct 2025 17:00:00 -0700
From: Dave Jiang <dave.jiang@...el.com>
To: "Fabio M. De Francesco" <fabio.m.de.francesco@...ux.intel.com>,
 linux-cxl@...r.kernel.org
Cc: Davidlohr Bueso <dave@...olabs.net>,
 Jonathan Cameron <jonathan.cameron@...wei.com>,
 Alison Schofield <alison.schofield@...el.com>,
 Vishal Verma <vishal.l.verma@...el.com>, Ira Weiny <ira.weiny@...el.com>,
 Dan Williams <dan.j.williams@...el.com>, Jonathan Corbet <corbet@....net>,
 linux-doc@...r.kernel.org, linux-kernel@...r.kernel.org,
 Gregory Price <gourry@...rry.net>, Robert Richter <rrichter@....com>,
 Cheatham Benjamin <benjamin.cheatham@....com>
Subject: Re: [PATCH 2/4 v5] cxl/core: Add helpers to detect Low Memory Holes
 on x86



On 10/6/25 8:58 AM, Fabio M. De Francesco wrote:
> On a x86 platform with a low memory hole (LHM), the BIOS may publish
> CFMWS that describes a system physical address (SPA) range that
> typically is only a subset of the corresponding CXL intermediate switch
> and endpoint decoder's host physical address (HPA) ranges. The CFMWS
> range never intersects the LHM and so the driver instantiates a root
> decoder whose HPA range size doesn't fully contain the matching switch
> and endpoint decoders' HPA ranges.[1]
> 
> To construct regions and attach decoders, the driver needs to match root
> decoders and regions with endpoint decoders. The process fails and
> returns errors because the driver is not designed to deal with SPA
> ranges which are smaller than the corresponding hardware decoders HPA
> ranges.
> 
> Introduce two functions that indirectly detect the presence of x86 LMH
> and allow the matching between a root decoder or an already constructed
> region with a corresponding intermediate switch or endpoint decoder to
> enable the construction of a region and the subsequent attachment of the
> same decoders to that region.
> 
> These functions return true when SPA/HPA misalignments due to LMH's are
> detected under specific conditions:
> 
> - Both the SPA and HPA ranges must start at LMH_CFMWS_RANGE_START (i.e.,
>   0x0 on x86 with LMH's).
> - The SPA range's size is less than HPA's.
> - The SPA range's size is less than 4G.
> - The HPA range's size is aligned to the NIW * 256M rule.
> 
> Also introduce a function that adjusts the range end of a region to be
> constructed and the DPA range's end of the endpoint decoders that will
> be later attached to that region.
> 
> [1] commit 7a81173f3 ("cxl: Documentation/driver-api/cxl: Describe the x86 Low Memory Hole solution")

I don't think this is the right hash.
c5dca38633da ("cxl: Documentation/driver-api/cxl: Describe the x86 Low Memory Hole solution")

> 
> Cc: Alison Schofield <alison.schofield@...el.com>
> Cc: Dan Williams <dan.j.williams@...el.com>
> Cc: Dave Jiang <dave.jiang@...el.com>
> Cc: Ira Weiny <ira.weiny@...el.com>
> Signed-off-by: Fabio M. De Francesco <fabio.m.de.francesco@...ux.intel.com>
> ---
>  drivers/cxl/Kconfig                |  4 ++
>  drivers/cxl/core/Makefile          |  1 +
>  drivers/cxl/core/platform_quirks.c | 99 ++++++++++++++++++++++++++++++
>  drivers/cxl/core/platform_quirks.h | 33 ++++++++++
>  4 files changed, 137 insertions(+)
>  create mode 100644 drivers/cxl/core/platform_quirks.c
>  create mode 100644 drivers/cxl/core/platform_quirks.h
> 
> diff --git a/drivers/cxl/Kconfig b/drivers/cxl/Kconfig
> index 48b7314afdb8..03c0583bc9a3 100644
> --- a/drivers/cxl/Kconfig
> +++ b/drivers/cxl/Kconfig
> @@ -211,6 +211,10 @@ config CXL_REGION
>  
>  	  If unsure say 'y'
>  
> +config CXL_PLATFORM_QUIRKS
> +	def_bool y
> +	depends on CXL_REGION
> +
>  config CXL_REGION_INVALIDATION_TEST
>  	bool "CXL: Region Cache Management Bypass (TEST)"
>  	depends on CXL_REGION
> diff --git a/drivers/cxl/core/Makefile b/drivers/cxl/core/Makefile
> index 5ad8fef210b5..1684e46b8709 100644
> --- a/drivers/cxl/core/Makefile
> +++ b/drivers/cxl/core/Makefile
> @@ -17,6 +17,7 @@ cxl_core-y += cdat.o
>  cxl_core-y += ras.o
>  cxl_core-$(CONFIG_TRACING) += trace.o
>  cxl_core-$(CONFIG_CXL_REGION) += region.o
> +cxl_core-$(CONFIG_CXL_PLATFORM_QUIRKS) += platform_quirks.o
>  cxl_core-$(CONFIG_CXL_MCE) += mce.o
>  cxl_core-$(CONFIG_CXL_FEATURES) += features.o
>  cxl_core-$(CONFIG_CXL_EDAC_MEM_FEATURES) += edac.o
> diff --git a/drivers/cxl/core/platform_quirks.c b/drivers/cxl/core/platform_quirks.c
> new file mode 100644
> index 000000000000..7e76e392b1ae
> --- /dev/null
> +++ b/drivers/cxl/core/platform_quirks.c
> @@ -0,0 +1,99 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +
> +#include <linux/range.h>
> +#include "platform_quirks.h"
> +#include "cxlmem.h"
> +#include "core.h"
> +
> +/* Start of CFMWS range that end before x86 Low Memory Holes */
> +#define LMH_CFMWS_RANGE_START 0x0ULL
> +
> +/**
> + * platform_cxlrd_matches_cxled() - Platform quirk to match CXL Root and
> + * Endpoint Decoders. It allows matching on platforms with LMH's.
> + * @cxlrd: The Root Decoder against which @cxled is tested for matching.
> + * @cxled: The Endpoint Decoder to be tested for matching @cxlrd.
> + *
> + * platform_cxlrd_matches_cxled() is typically called from the
> + * match_*_by_range() functions in region.c. It checks if an endpoint decoder
> + * matches a given root decoder and returns true to allow the driver to succeed
> + * in the construction of regions where it would otherwise fail for the presence
> + * of a Low Memory Hole (see Documentation/driver-api/cxl/conventions.rst).
> + *
> + * In x86 platforms with LMH's, the CFMWS ranges never intersect the LMH, the
> + * endpoint decoder's HPA range size is always guaranteed aligned to NIW*256MB
> + * and also typically larger than the matching root decoder's, and the root
> + * decoder's range end is at an address that is necessarily less than SZ_4G
> + * (i.e., the Hole is in Low Memory - this function doesn't deal with other
> + * kinds of holes).
> + *
> + * Return: true if an endpoint matches a root decoder, else false.
> + */
> +bool platform_cxlrd_matches_cxled(const struct cxl_root_decoder *cxlrd,
> +				  const struct cxl_endpoint_decoder *cxled)
> +{
> +	const struct range *rd_r, *sd_r;
> +	int align;
> +
> +	rd_r = &cxlrd->cxlsd.cxld.hpa_range;
> +	sd_r = &cxled->cxld.hpa_range;
> +	align = cxled->cxld.interleave_ways * SZ_256M;
> +
> +	if (rd_r->start == LMH_CFMWS_RANGE_START &&
> +	    rd_r->start == sd_r->start && rd_r->end < sd_r->end &&

Break this into 2 lines to make it more readable

> +	    rd_r->end < (LMH_CFMWS_RANGE_START + SZ_4G) &&
> +	    IS_ALIGNED(range_len(sd_r), align))
> +		return true;
> +
> +	return false;
> +}
> +
> +/**
> + * platform_region_matches_cxld() - Platform quirk to match a CXL Region and a
> + * Switch or Endpoint Decoder. It allows matching on platforms with LMH's.
> + * @p: Region Params against which @cxled is matched.
> + * @cxld: Switch or Endpoint Decoder to be tested for matching @p.
> + *
> + * Similar to platform_cxlrd_matches_cxled(), it matches regions and
> + * decoders on platforms with LMH's.
> + *
> + * Return: true if a Decoder matches a Region, else false.
> + */
> +bool platform_region_matches_cxld(const struct cxl_region_params *p,
> +				  const struct cxl_decoder *cxld)
> +{
> +	const struct range *r = &cxld->hpa_range;
> +	const struct resource *res = p->res;
> +	int align = cxld->interleave_ways * SZ_256M;
> +
> +	if (res->start == LMH_CFMWS_RANGE_START && res->start == r->start &&
> +	    res->end < r->end && res->end < (LMH_CFMWS_RANGE_START + SZ_4G) &&

Break the first and this line into 2 lines each with a single compare to make it more readable

> +	    IS_ALIGNED(range_len(r), align))
> +		return true;
> +
> +	return false;
> +}
> +
> +void platform_res_adjust(struct resource *res,
> +			 struct cxl_endpoint_decoder *cxled,
> +			 const struct cxl_root_decoder *cxlrd)
> +{
> +	if (!platform_cxlrd_matches_cxled(cxlrd, cxled))
> +		return;
> +
> +	guard(rwsem_write)(&cxl_rwsem.dpa);
> +	dev_dbg(cxled_to_memdev(cxled)->dev.parent,
> +		"Low Memory Hole detected. Resources were (%s: %pr, %pr)\n",
> +		dev_name(&cxled->cxld.dev), res, cxled->dpa_res);
> +	if (res) {
> +		/* Trim region resource overlap with LMH */
> +		res->end = cxlrd->res->end;
> +	}
> +	/* Match endpoint decoder's DPA resource to root decoder's */
> +	cxled->dpa_res->end =
> +		cxled->dpa_res->start +

this can be on the same line as the first

DJ

> +		resource_size(cxlrd->res) / cxled->cxld.interleave_ways - 1;
> +	dev_info(cxled_to_memdev(cxled)->dev.parent,
> +		 "Resources have been adjusted for LMH (%s: %pr, %pr)\n",
> +		 dev_name(&cxled->cxld.dev), res, cxled->dpa_res);
> +}
> diff --git a/drivers/cxl/core/platform_quirks.h b/drivers/cxl/core/platform_quirks.h
> new file mode 100644
> index 000000000000..a15592b4e90e
> --- /dev/null
> +++ b/drivers/cxl/core/platform_quirks.h
> @@ -0,0 +1,33 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +
> +#include "cxl.h"
> +
> +#ifdef CONFIG_CXL_PLATFORM_QUIRKS
> +bool platform_cxlrd_matches_cxled(const struct cxl_root_decoder *cxlrd,
> +				  const struct cxl_endpoint_decoder *cxled);
> +bool platform_region_matches_cxld(const struct cxl_region_params *p,
> +				  const struct cxl_decoder *cxld);
> +void platform_res_adjust(struct resource *res,
> +			 struct cxl_endpoint_decoder *cxled,
> +			 const struct cxl_root_decoder *cxlrd);
> +#else
> +static inline bool
> +platform_root_decoder_contains(const struct cxl_root_decoder *cxlrd,
> +			       const struct cxl_endpoint_decoder *cxled)
> +{
> +	return false;
> +}
> +
> +static inline bool
> +platform_region_matches_cxld(const struct cxl_region_params *p,
> +			     const struct cxl_decoder *cxld)
> +{
> +	return false;
> +}
> +
> +inline void platform_res_adjust(struct resource *res,
> +				struct cxl_endpoint_decoder *cxled,
> +				const struct cxl_root_decoder *cxlrd)
> +{
> +}
> +#endif /* CONFIG_CXL_PLATFORM_QUIRKS */

Powered by blists - more mailing lists