[PATCH v2 04/15] drm/i915: Bypass LMEMBAR/GTTMMADR for MTL stolen memory access

[Nirmoy Das]

Hi Ville,

Apologies, but I lost track of this series after I returned from sick leave.


On 12/15/2023 11:59 AM, Ville Syrjala wrote:
> From: Ville Syrjälä <ville.syrjala at linux.intel.com>
>
> On MTL accessing stolen memory via the BARs is somehow borked,
> and it can hang the machine. As a workaround let's bypass the
> BARs and just go straight to DSMBASE/GSMBASE instead.
>
> Note that on every other platform this itself would hang the
> machine, but on MTL the system firmware is expected to relax
> the access permission guarding stolen memory to enable this
> workaround, and thus direct CPU accesses should be fine.
>
> TODO: add w/a numbers and whatnot
>
> Cc: Paz Zcharya <pazz at chromium.org>
> Cc: Nirmoy Das <nirmoy.das at intel.com>
> Cc: Radhakrishna Sripada <radhakrishna.sripada at intel.com>
> Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> Signed-off-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> ---
>   drivers/gpu/drm/i915/gem/i915_gem_stolen.c | 11 ++++++++++-
>   drivers/gpu/drm/i915/gt/intel_ggtt.c       | 13 ++++++++++++-
>   2 files changed, 22 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_stolen.c b/drivers/gpu/drm/i915/gem/i915_gem_stolen.c
> index ee237043c302..252fe5cd6ede 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_stolen.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_stolen.c
> @@ -941,7 +941,16 @@ i915_gem_stolen_lmem_setup(struct drm_i915_private *i915, u16 type,
>   		dsm_size = ALIGN_DOWN(lmem_size - dsm_base, SZ_1M);
>   	}
>   
> -	if (pci_resource_len(pdev, GEN12_LMEM_BAR) < lmem_size) {
> +	if (IS_METEORLAKE(i915)) {
> +		/*
> +		 * Workaround: access via BAR can hang MTL, go directly to DSM.
> +		 *
> +		 * Normally this would not work but on MTL the system firmware
> +		 * should have relaxed the access permissions sufficiently.
> +		 */
> +		io_start = intel_uncore_read64(uncore, GEN12_DSMBASE) & GEN12_BDSM_MASK;
> +		io_size = dsm_size;

This will work well on host driver but I am afraid this will not work on 
VM when someone tries to do direct device assignment of the igfx.

GSMBASE/DSMBASE is reserved region so won't show up in VM, last I checked.

This is an obscure usages but are we suppose to support that? If so then 
we need to detect that and fall back to binder approach.


Regards,

Nirmoy

> +	} else if (pci_resource_len(pdev, GEN12_LMEM_BAR) < lmem_size) {
>   		io_start = 0;
>   		io_size = 0;
>   	} else {
> diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c
> index 21a7e3191c18..ab71d74ec426 100644
> --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c
> +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c
> @@ -24,6 +24,7 @@
>   #include "intel_ring.h"
>   #include "i915_drv.h"
>   #include "i915_pci.h"
> +#include "i915_reg.h"
>   #include "i915_request.h"
>   #include "i915_scatterlist.h"
>   #include "i915_utils.h"
> @@ -1152,13 +1153,23 @@ static unsigned int gen6_gttadr_offset(struct drm_i915_private *i915)
>   static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size)
>   {
>   	struct drm_i915_private *i915 = ggtt->vm.i915;
> +	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
>   	struct pci_dev *pdev = to_pci_dev(i915->drm.dev);
>   	phys_addr_t phys_addr;
>   	u32 pte_flags;
>   	int ret;
>   
>   	GEM_WARN_ON(pci_resource_len(pdev, GEN4_GTTMMADR_BAR) != gen6_gttmmadr_size(i915));
> -	phys_addr = pci_resource_start(pdev, GEN4_GTTMMADR_BAR) + gen6_gttadr_offset(i915);
> +	/*
> +	 * Workaround: access via BAR can hang MTL, go directly to GSM.
> +	 *
> +	 * Normally this would not work but on MTL the system firmware
> +	 * should have relaxed the access permissions sufficiently.
> +	 */
> +	if (IS_METEORLAKE(i915))
> +		phys_addr = intel_uncore_read64(uncore, GEN12_GSMBASE) & GEN12_BDSM_MASK;
> +	else
> +		phys_addr = pci_resource_start(pdev, GEN4_GTTMMADR_BAR) + gen6_gttadr_offset(i915);
>   
>   	if (needs_wc_ggtt_mapping(i915))
>   		ggtt->gsm = ioremap_wc(phys_addr, size);