[PATCH v3] drm/i915/mtl: enable local stolen memory

Matt Roper matthew.d.roper at intel.com
Tue Sep 27 22:22:02 UTC 2022


On Tue, Sep 27, 2022 at 12:14:24AM +0530, Aravind Iddamsetty wrote:
> As an integrated GPU, MTL does not have local memory and
> HAS_LMEM() returns false.  However the platform's stolen memory
> is presented via BAR2 (i.e., the BAR we traditionally consider
> to be the LMEM BAR) and should be managed by the driver the same
> way that local memory is on dgpu platforms (which includes
> setting the "lmem" bit on page table entries).  We use the term
> "local stolen memory" to refer to this model.
> 
> v2:
> 1. dropped is_dsm_invalid, updated valid_stolen_size check from Lucas
> (Jani, Lucas)
> 2. drop lmembar_is_igpu_stolen
> 3. revert to referring GFXMEM_BAR as GEN12_LMEM_BAR (Lucas)
> 
> v3:(Jani)
> 1. rename get_mtl_gms_size to mtl_get_gms_size
> 2. define register for MMIO address
> 
> Cc: Matt Roper <matthew.d.roper at intel.com>
> Cc: Lucas De Marchi <lucas.demarchi at intel.com>
> Cc: Jani Nikula <jani.nikula at linux.intel.com>
> 

Since this stuff is somewhat hard to find documentation on, you might
want to include a bspec page number or two here.

Bspec: 63830

seems like a useful one for reference at least.

> Signed-off-by: CQ Tang <cq.tang at intel.com>
> Signed-off-by: Aravind Iddamsetty <aravind.iddamsetty at intel.com>
> Original-author: CQ Tang
> ---
>  drivers/gpu/drm/i915/gem/i915_gem_stolen.c | 88 ++++++++++++++++++----
>  drivers/gpu/drm/i915/gt/intel_ggtt.c       |  2 +-
>  drivers/gpu/drm/i915/i915_drv.h            |  3 +
>  drivers/gpu/drm/i915/i915_reg.h            |  5 ++
>  4 files changed, 81 insertions(+), 17 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_stolen.c b/drivers/gpu/drm/i915/gem/i915_gem_stolen.c
> index c5a4035c99cd..0eb66c55bbf3 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_stolen.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_stolen.c
> @@ -77,9 +77,9 @@ void i915_gem_stolen_remove_node(struct drm_i915_private *i915,
>  	mutex_unlock(&i915->mm.stolen_lock);
>  }
>  
> -static bool valid_stolen_size(struct resource *dsm)
> +static bool valid_stolen_size(struct drm_i915_private *i915, struct resource *dsm)
>  {
> -	return dsm->start != 0 && dsm->end > dsm->start;
> +	return (dsm->start != 0 || HAS_BAR2_SMEM_STOLEN(i915)) && dsm->end > dsm->start;
>  }
>  
>  static int adjust_stolen(struct drm_i915_private *i915,
> @@ -88,7 +88,7 @@ static int adjust_stolen(struct drm_i915_private *i915,
>  	struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
>  	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
>  
> -	if (!valid_stolen_size(dsm))
> +	if (!valid_stolen_size(i915, dsm))
>  		return -EINVAL;
>  
>  	/*
> @@ -135,7 +135,7 @@ static int adjust_stolen(struct drm_i915_private *i915,
>  		}
>  	}
>  
> -	if (!valid_stolen_size(dsm))
> +	if (!valid_stolen_size(i915, dsm))
>  		return -EINVAL;
>  
>  	return 0;
> @@ -148,9 +148,10 @@ static int request_smem_stolen(struct drm_i915_private *i915,
>  
>  	/*
>  	 * With stolen lmem, we don't need to request system memory for the
> -	 * address range since it's local to the gpu.
> +	 * address range since it's local to the gpu and in some IGFX devices
> +	 * BAR2 is exposed as stolen

Minor nitpick:  this sentence is a bit hard to read/understand.  I'd
leave the original sentence as is and add a separate sentence explaining
the situation for igpu platforms with stolen memory exposed through
BAR2.

>  	 */
> -	if (HAS_LMEM(i915))
> +	if (HAS_LMEM(i915) || HAS_BAR2_SMEM_STOLEN(i915))
>  		return 0;
>  
>  	/*
> @@ -385,8 +386,6 @@ static void icl_get_stolen_reserved(struct drm_i915_private *i915,
>  
>  	drm_dbg(&i915->drm, "GEN6_STOLEN_RESERVED = 0x%016llx\n", reg_val);
>  
> -	*base = reg_val & GEN11_STOLEN_RESERVED_ADDR_MASK;
> -
>  	switch (reg_val & GEN8_STOLEN_RESERVED_SIZE_MASK) {
>  	case GEN8_STOLEN_RESERVED_1M:
>  		*size = 1024 * 1024;
> @@ -404,6 +403,12 @@ static void icl_get_stolen_reserved(struct drm_i915_private *i915,
>  		*size = 8 * 1024 * 1024;
>  		MISSING_CASE(reg_val & GEN8_STOLEN_RESERVED_SIZE_MASK);
>  	}
> +
> +	if (HAS_BAR2_SMEM_STOLEN(i915))
> +		/* the base is initialized to stolen top so subtract size to get base */
> +		*base -= *size;
> +	else
> +		*base = reg_val & GEN11_STOLEN_RESERVED_ADDR_MASK;
>  }
>  
>  /*
> @@ -833,6 +838,34 @@ static const struct intel_memory_region_ops i915_region_stolen_lmem_ops = {
>  	.init_object = _i915_gem_object_stolen_init,
>  };
>  
> +static int mtl_get_gms_size(struct intel_uncore *uncore)
> +{
> +	u16 ggc, gms;
> +
> +	ggc = intel_uncore_read16(uncore, GGC);
> +
> +	/* check GGMS, should be fixed 0x3 (8MB) */
> +	if ((ggc & GGMS_MASK) != GGMS_MASK)
> +		return -EIO;
> +
> +	/* return valid GMS value, -EIO if invalid */
> +	gms = (ggc & GMS_MASK) >> GMS_SHIFT;

It's better to write this as

        gms = REG_FIELD_GET(GMS_MASK, ggc);

You can also eliminate the definition of GMS_SHIFT with that.

> +	switch (gms) {
> +	case 0x0 ... 0x10:

Shouldn't this just be 0x0 ... 0x4?

> +		return gms * 32;
> +	case 0x20:
> +		return 1024;
> +	case 0x30:
> +		return 1536;
> +	case 0x40:
> +		return 2048;

I don't see 0x20, 0x30, or 0x40 listed as valid values in the bspec
anymore.


> +	case 0xf0 ... 0xfe:
> +		return (gms - 0xf0 + 1) * 4;
> +	default:

You might want to add a MISSING_CASE(gms) since this isn't supposed to
be able to happen, but if it does, we'd like to see which invalid value
we actually received.

> +		return -EIO;
> +	}
> +}
> +
>  struct intel_memory_region *
>  i915_gem_stolen_lmem_setup(struct drm_i915_private *i915, u16 type,
>  			   u16 instance)
> @@ -843,6 +876,7 @@ i915_gem_stolen_lmem_setup(struct drm_i915_private *i915, u16 type,
>  	struct intel_memory_region *mem;
>  	resource_size_t io_start, io_size;
>  	resource_size_t min_page_size;
> +	int ret;
>  
>  	if (WARN_ON_ONCE(instance))
>  		return ERR_PTR(-ENODEV);
> @@ -850,12 +884,8 @@ i915_gem_stolen_lmem_setup(struct drm_i915_private *i915, u16 type,
>  	if (!i915_pci_resource_valid(pdev, GEN12_LMEM_BAR))
>  		return ERR_PTR(-ENXIO);
>  
> -	/* Use DSM base address instead for stolen memory */
> -	dsm_base = intel_uncore_read64(uncore, GEN12_DSMBASE) & GEN12_BDSM_MASK;
> -	if (IS_DG1(uncore->i915)) {
> +	if (HAS_BAR2_SMEM_STOLEN(i915) || IS_DG1(i915)) {
>  		lmem_size = pci_resource_len(pdev, GEN12_LMEM_BAR);
> -		if (WARN_ON(lmem_size < dsm_base))
> -			return ERR_PTR(-ENODEV);
>  	} else {
>  		resource_size_t lmem_range;
>  
> @@ -864,13 +894,39 @@ i915_gem_stolen_lmem_setup(struct drm_i915_private *i915, u16 type,
>  		lmem_size *= SZ_1G;
>  	}
>  
> -	dsm_size = lmem_size - dsm_base;
> -	if (pci_resource_len(pdev, GEN12_LMEM_BAR) < lmem_size) {
> +	if (HAS_BAR2_SMEM_STOLEN(i915)) {
> +		/*
> +		 * MTL dsm size is in GGC register, not the bar size.
> +		 * also MTL uses offset to DSMBASE in ptes, so i915
> +		 * uses dsm_base = 0 to setup stolen region.
> +		 */
> +		ret = mtl_get_gms_size(uncore);
> +		if (ret < 0) {
> +			drm_err(&i915->drm, "invalid MTL GGC register setting\n");
> +			return ERR_PTR(ret);
> +		}
> +
> +		dsm_base = 0;
> +		dsm_size = (resource_size_t)(ret * SZ_1M);
> +
> +		GEM_BUG_ON(pci_resource_len(pdev, GEN12_LMEM_BAR) != 256 * SZ_1M);

You could simplify to just SZ_256M here.  Also SZ_8M for some of the
other spots below.


Matt

> +		GEM_BUG_ON((dsm_size + 8 * SZ_1M) > lmem_size);
> +	} else {
> +		/* Use DSM base address instead for stolen memory */
> +		dsm_base = intel_uncore_read64(uncore, GEN12_DSMBASE);
> +		if (WARN_ON(lmem_size < dsm_base))
> +			return ERR_PTR(-ENODEV);
> +		dsm_size = lmem_size - dsm_base;
> +	}
> +
> +	io_size = dsm_size;
> +	if (pci_resource_len(pdev, GEN12_LMEM_BAR) < dsm_size) {
>  		io_start = 0;
>  		io_size = 0;
> +	} else if (HAS_BAR2_SMEM_STOLEN(i915)) {
> +		io_start = pci_resource_start(pdev, GEN12_LMEM_BAR) + 8 * SZ_1M;
>  	} else {
>  		io_start = pci_resource_start(pdev, GEN12_LMEM_BAR) + dsm_base;
> -		io_size = dsm_size;
>  	}
>  
>  	min_page_size = HAS_64K_PAGES(i915) ? I915_GTT_PAGE_SIZE_64K :
> diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c
> index 30cf5c3369d9..b31fe0fb013f 100644
> --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c
> +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c
> @@ -931,7 +931,7 @@ static int gen8_gmch_probe(struct i915_ggtt *ggtt)
>  	unsigned int size;
>  	u16 snb_gmch_ctl;
>  
> -	if (!HAS_LMEM(i915)) {
> +	if (!HAS_LMEM(i915) && !HAS_BAR2_SMEM_STOLEN(i915)) {
>  		if (!i915_pci_resource_valid(pdev, GTT_APERTURE_BAR))
>  			return -ENXIO;
>  
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index d830d52ded5d..b33ba0d49bbd 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -975,6 +975,9 @@ IS_SUBPLATFORM(const struct drm_i915_private *i915,
>  
>  #define HAS_ONE_EU_PER_FUSE_BIT(i915)	(INTEL_INFO(i915)->has_one_eu_per_fuse_bit)
>  
> +#define HAS_BAR2_SMEM_STOLEN(i915) (!HAS_LMEM(i915) && \
> +				    GRAPHICS_VER_FULL(i915) >= IP_VER(12, 70))
> +
>  /* intel_device_info.c */
>  static inline struct intel_device_info *
>  mkwrite_device_info(struct drm_i915_private *dev_priv)
> diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h
> index 5003a5ffbc6a..3ace2d6b4961 100644
> --- a/drivers/gpu/drm/i915/i915_reg.h
> +++ b/drivers/gpu/drm/i915/i915_reg.h
> @@ -7967,6 +7967,11 @@ enum skl_power_gate {
>  							   _ICL_PIPE_DSS_CTL2_PB, \
>  							   _ICL_PIPE_DSS_CTL2_PC)
>  
> +#define GGC				_MMIO(0x108040)
> +#define   GMS_MASK			REG_GENMASK(15, 8)
> +#define   GMS_SHIFT			8
> +#define   GGMS_MASK			REG_GENMASK(7, 6)
> +
>  #define GEN12_GSMBASE			_MMIO(0x108100)
>  #define GEN12_DSMBASE			_MMIO(0x1080C0)
>  #define   GEN12_BDSM_MASK		REG_GENMASK64(63, 20)
> -- 
> 2.25.1
> 

-- 
Matt Roper
Graphics Software Engineer
VTT-OSGC Platform Enablement
Intel Corporation


More information about the dri-devel mailing list