diff mbox series

[5/5] drm/msm/a6xx: Add support for using system cache(LLC)

Message ID 1576761286-20451-6-git-send-email-smasetty@codeaurora.org (mailing list archive)
State Superseded
Headers show
Series drm/msm/a6xx: System Cache Support | expand

Commit Message

Sharat Masetty Dec. 19, 2019, 1:14 p.m. UTC
The last level system cache can be partitioned to 32 different slices
of which GPU has two slices preallocated. One slice is used for caching GPU
buffers and the other slice is used for caching the GPU SMMU pagetables.
This patch talks to the core system cache driver to acquire the slice handles,
configure the SCID's to those slices and activates and deactivates the slices
upon GPU power collapse and restore.

Some support from the IOMMU driver is also needed to make use of the
system cache. IOMMU_QCOM_SYS_CACHE is a buffer protection flag which enables
caching GPU data buffers in the system cache with memory attributes such
as outer cacheable, read-allocate, write-allocate for buffers. The GPU
then has the ability to override a few cacheability parameters which it
does to override write-allocate to write-no-allocate as the GPU hardware
does not benefit much from it.

Similarly DOMAIN_ATTR_QCOM_SYS_CACHE is another domain level attribute
used by the IOMMU driver to set the right attributes to cache the hardware
pagetables into the system cache.

Signed-off-by: Sharat Masetty <smasetty@codeaurora.org>
---
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 122 +++++++++++++++++++++++++++++++++-
 drivers/gpu/drm/msm/adreno/a6xx_gpu.h |   9 +++
 drivers/gpu/drm/msm/msm_iommu.c       |  13 ++++
 drivers/gpu/drm/msm/msm_mmu.h         |   3 +
 4 files changed, 146 insertions(+), 1 deletion(-)

--
1.9.1

Comments

Jordan Crouse Dec. 19, 2019, 7:58 p.m. UTC | #1
On Thu, Dec 19, 2019 at 06:44:46PM +0530, Sharat Masetty wrote:
> The last level system cache can be partitioned to 32 different slices
> of which GPU has two slices preallocated. One slice is used for caching GPU
> buffers and the other slice is used for caching the GPU SMMU pagetables.
> This patch talks to the core system cache driver to acquire the slice handles,
> configure the SCID's to those slices and activates and deactivates the slices
> upon GPU power collapse and restore.
> 
> Some support from the IOMMU driver is also needed to make use of the
> system cache. IOMMU_QCOM_SYS_CACHE is a buffer protection flag which enables
> caching GPU data buffers in the system cache with memory attributes such
> as outer cacheable, read-allocate, write-allocate for buffers. The GPU
> then has the ability to override a few cacheability parameters which it
> does to override write-allocate to write-no-allocate as the GPU hardware
> does not benefit much from it.
> 
> Similarly DOMAIN_ATTR_QCOM_SYS_CACHE is another domain level attribute
> used by the IOMMU driver to set the right attributes to cache the hardware
> pagetables into the system cache.
> 
> Signed-off-by: Sharat Masetty <smasetty@codeaurora.org>
> ---
>  drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 122 +++++++++++++++++++++++++++++++++-
>  drivers/gpu/drm/msm/adreno/a6xx_gpu.h |   9 +++
>  drivers/gpu/drm/msm/msm_iommu.c       |  13 ++++
>  drivers/gpu/drm/msm/msm_mmu.h         |   3 +
>  4 files changed, 146 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> index faff6ff..0c7fdee 100644
> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> @@ -9,6 +9,7 @@
>  #include "a6xx_gmu.xml.h"
> 
>  #include <linux/devfreq.h>
> +#include <linux/soc/qcom/llcc-qcom.h>
> 
>  #define GPU_PAS_ID 13
> 
> @@ -781,6 +782,117 @@ static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu)
>  	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
>  }
> 
> +#define A6XX_LLC_NUM_GPU_SCIDS		5
> +#define A6XX_GPU_LLC_SCID_NUM_BITS	5

As I mention below, I'm not sure if we need these 

> +#define A6XX_GPU_LLC_SCID_MASK \
> +	((1 << (A6XX_LLC_NUM_GPU_SCIDS * A6XX_GPU_LLC_SCID_NUM_BITS)) - 1)
> +
> +#define A6XX_GPUHTW_LLC_SCID_SHIFT	25
> +#define A6XX_GPUHTW_LLC_SCID_MASK \
> +	(((1 << A6XX_GPU_LLC_SCID_NUM_BITS) - 1) << A6XX_GPUHTW_LLC_SCID_SHIFT)
> +

Normally these go into the envytools regmap but if we're going to do these guys
lets use the power of <linux/bitfield.h> for good.

#define A6XX_GPU_LLC_SCID GENMASK(24, 0)
#define A6XX_GPUHTW_LLC_SCID GENMASK(29, 25)

> +static inline void a6xx_gpu_cx_rmw(struct a6xx_llc *llc,

Don't mark C functions as inline - let the compiler figure it out for you.

> +	u32 reg, u32 mask, u32 or)
> +{
> +	msm_rmw(llc->mmio + (reg << 2), mask, or);
> +}
> +
> +static void a6xx_llc_deactivate(struct a6xx_llc *llc)
> +{
> +	llcc_slice_deactivate(llc->gpu_llc_slice);
> +	llcc_slice_deactivate(llc->gpuhtw_llc_slice);
> +}
> +
> +static void a6xx_llc_activate(struct a6xx_llc *llc)
> +{
> +	if (!llc->mmio)
> +		return;
> +
> +	/* Program the sub-cache ID for all GPU blocks */
> +	if (!llcc_slice_activate(llc->gpu_llc_slice))
> +		a6xx_gpu_cx_rmw(llc,
> +				REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_1,
> +				A6XX_GPU_LLC_SCID_MASK,
> +				(llc->cntl1_regval &
> +				 A6XX_GPU_LLC_SCID_MASK));

This is out of order with the comments below, but if we store the slice id then
you could calculate regval here and not have to store it.

> +
> +	/* Program the sub-cache ID for the GPU pagetables */
> +	if (!llcc_slice_activate(llc->gpuhtw_llc_slice))

val |= FIELD_SET(A6XX_GPUHTW_LLC_SCID, htw_llc_sliceid);

> +		a6xx_gpu_cx_rmw(llc,
> +				REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_1,
> +				A6XX_GPUHTW_LLC_SCID_MASK,
> +				(llc->cntl1_regval &
> +				 A6XX_GPUHTW_LLC_SCID_MASK));

And this could be FIELD_SET(A6XX_GPUHTW_LLC_SCID, sliceid);

In theory you could just calculate the u32 and write it directly without a rmw.
In fact, that might be preferable - if the slice activate failed, you don't want
to run the risk that the scid for htw is still populated.

> +
> +	/* Program cacheability overrides */
> +	a6xx_gpu_cx_rmw(llc, REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_0, 0xF,
> +		llc->cntl0_regval);

As below, this could easily be a constant.

> +}
> +
> +static void a6xx_llc_slices_destroy(struct a6xx_llc *llc)
> +{
> +	if (llc->mmio)
> +		iounmap(llc->mmio);

msm_ioremap returns a devm_ managed resource, so do not use iounmap() to free
it. Bets to just leave it and let the gpu device handle it when it goes boom.

> +
> +	llcc_slice_putd(llc->gpu_llc_slice);
> +	llcc_slice_putd(llc->gpuhtw_llc_slice);
> +}
> +
> +static int a6xx_llc_slices_init(struct platform_device *pdev,
T
This can be void, I don't think we care if it passes or fails.

> +		struct a6xx_llc *llc)
> +{
> +	llc->mmio = msm_ioremap(pdev, "cx_mem", "gpu_cx");
> +	if (IS_ERR_OR_NULL(llc->mmio))

msm_ioremap can not return NULL.

> +		return -ENODEV;
> +
> +	llc->gpu_llc_slice = llcc_slice_getd(LLCC_GPU);
> +	llc->gpuhtw_llc_slice = llcc_slice_getd(LLCC_GPUHTW);
> +	if (IS_ERR(llc->gpu_llc_slice) && IS_ERR(llc->gpuhtw_llc_slice))
> +		return -ENODEV;
> +
> +	/*
> +	 * CNTL0 provides options to override the settings for the
> +	 * read and write allocation policies for the LLC. These
> +	 * overrides are global for all memory transactions from
> +	 * the GPU.
> +	 *
> +	 * 0x3: read-no-alloc-overridden = 0
> +	 *      read-no-alloc = 0 - Allocate lines on read miss
> +	 *      write-no-alloc-overridden = 1
> +	 *      write-no-alloc = 1 - Do not allocates lines on write miss
> +	 */
> +	llc->cntl0_regval = 0x03;

This is a fixed value isn't it?  We should be able to get away with writing a
constant.

> +
> +	/*
> +	 * CNTL1 is used to specify SCID for (CP, TP, VFD, CCU and UBWC
> +	 * FLAG cache) GPU blocks. This value will be passed along with
> +	 * the address for any memory transaction from GPU to identify
> +	 * the sub-cache for that transaction.
> +	 */
> +	if (!IS_ERR(llc->gpu_llc_slice)) {
> +		u32 gpu_scid = llcc_get_slice_id(llc->gpu_llc_slice);
> +		int i;
> +
> +		for (i = 0; i < A6XX_LLC_NUM_GPU_SCIDS; i++)
> +			llc->cntl1_regval |=
> +				gpu_scid << (A6XX_GPU_LLC_SCID_NUM_BITS * i);

As above, i'm not sure a loop is better than just:

gpu_scid &= 0x1f;

llc->cntl1_regval = (gpu_scid << 0) || (gpu_scid << 5) | (gpu_scid << 10)
 | (gpu_scid << 15) | (gpu_scid << 20);

And I'm not even sure we need do this math here in the first place.

> +	}
> +
> +	/*
> +	 * Set SCID for GPU IOMMU. This will be used to access
> +	 * page tables that are cached in LLC.
> +	 */
> +	if (!IS_ERR(llc->gpuhtw_llc_slice)) {
> +		u32 gpuhtw_scid = llcc_get_slice_id(llc->gpuhtw_llc_slice);
> +
> +		llc->cntl1_regval |=
> +			gpuhtw_scid << A6XX_GPUHTW_LLC_SCID_SHIFT;
> +	}

As above, I think storing the slice id could be more beneficial than calculating
a value, but if we do calculate a value, use FIELD_SET(A6XX_GPUHTW_LLC_SCID, )

> +
> +	return 0;
> +}
> +
>  static int a6xx_pm_resume(struct msm_gpu *gpu)
>  {
>  	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> @@ -795,6 +907,8 @@ static int a6xx_pm_resume(struct msm_gpu *gpu)
> 
>  	msm_gpu_resume_devfreq(gpu);
> 
> +	a6xx_llc_activate(&a6xx_gpu->llc);
> +
>  	return 0;
>  }
> 
> @@ -803,6 +917,8 @@ static int a6xx_pm_suspend(struct msm_gpu *gpu)
>  	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
>  	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
> 
> +	a6xx_llc_deactivate(&a6xx_gpu->llc);
> +
>  	devfreq_suspend_device(gpu->devfreq.devfreq);
> 
>  	/*
> @@ -851,6 +967,7 @@ static void a6xx_destroy(struct msm_gpu *gpu)
>  		drm_gem_object_put_unlocked(a6xx_gpu->sqe_bo);
>  	}
> 
> +	a6xx_llc_slices_destroy(&a6xx_gpu->llc);
>  	a6xx_gmu_remove(a6xx_gpu);
> 
>  	adreno_gpu_cleanup(adreno_gpu);
> @@ -924,7 +1041,10 @@ struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
>  	adreno_gpu->registers = NULL;
>  	adreno_gpu->reg_offsets = a6xx_register_offsets;
> 
> -	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1, 0);
> +	ret = a6xx_llc_slices_init(pdev, &a6xx_gpu->llc);
> +

Confirming we don't care if a6xx_llc_slices_init passes or fails.

> +	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1,
> +			ret ? 0 : MMU_FEATURE_USE_SYSTEM_CACHE);
>  	if (ret) {
>  		a6xx_destroy(&(a6xx_gpu->base.base));
>  		return ERR_PTR(ret);
> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> index 7239b8b..09b9ad0 100644
> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
> @@ -12,6 +12,14 @@
> 
>  extern bool hang_debug;
> 
> +struct a6xx_llc {
> +	void __iomem *mmio;
> +	void *gpu_llc_slice;
> +	void *gpuhtw_llc_slice;
> +	u32 cntl0_regval;

As above, I'm not sure if cntl0 is needed.  Heck, I'm not even sure cntl1 is
needed - since we could store or query the ids at activate time.

> +	u32 cntl1_regval;
> +};
> +
>  struct a6xx_gpu {
>  	struct adreno_gpu base;
> 
> @@ -21,6 +29,7 @@ struct a6xx_gpu {
>  	struct msm_ringbuffer *cur_ring;
> 
>  	struct a6xx_gmu gmu;
> +	struct a6xx_llc llc;
>  };
> 
>  #define to_a6xx_gpu(x) container_of(x, struct a6xx_gpu, base)
> diff --git a/drivers/gpu/drm/msm/msm_iommu.c b/drivers/gpu/drm/msm/msm_iommu.c
> index 8c95c31..4699367 100644
> --- a/drivers/gpu/drm/msm/msm_iommu.c
> +++ b/drivers/gpu/drm/msm/msm_iommu.c
> @@ -27,6 +27,16 @@ static int msm_iommu_attach(struct msm_mmu *mmu, const char * const *names,
>  			    int cnt)
>  {
>  	struct msm_iommu *iommu = to_msm_iommu(mmu);
> +	int gpu_htw_llc = 1;
> +
> +	/*
> +	 * This allows GPU to set the bus attributes required
> +	 * to use system cache on behalf of the iommu page table
> +	 * walker.
> +	 */
> +	if (msm_mmu_has_feature(mmu, MMU_FEATURE_USE_SYSTEM_CACHE))
> +		iommu_domain_set_attr(iommu->domain,
> +				DOMAIN_ATTR_QCOM_SYS_CACHE, &gpu_htw_llc);

We're all okay if this fails?  No harm no foul?

> 
>  	return iommu_attach_device(iommu->domain, mmu->dev);
>  }
> @@ -45,6 +55,9 @@ static int msm_iommu_map(struct msm_mmu *mmu, uint64_t iova,
>  	struct msm_iommu *iommu = to_msm_iommu(mmu);
>  	size_t ret;
> 
> +	if (msm_mmu_has_feature(mmu, MMU_FEATURE_USE_SYSTEM_CACHE))
> +		prot |= IOMMU_QCOM_SYS_CACHE;
> +
>  	ret = iommu_map_sg(iommu->domain, iova, sgt->sgl, sgt->nents, prot);
>  	WARN_ON(!ret);
> 
> diff --git a/drivers/gpu/drm/msm/msm_mmu.h b/drivers/gpu/drm/msm/msm_mmu.h
> index 1e4ac36d..3e6bdad 100644
> --- a/drivers/gpu/drm/msm/msm_mmu.h
> +++ b/drivers/gpu/drm/msm/msm_mmu.h
> @@ -18,6 +18,9 @@ struct msm_mmu_funcs {
>  	void (*destroy)(struct msm_mmu *mmu);
>  };
> 
> +/* MMU features */
> +#define MMU_FEATURE_USE_SYSTEM_CACHE (1 << 0)
> +
>  struct msm_mmu {
>  	const struct msm_mmu_funcs *funcs;
>  	struct device *dev;
> --
> 1.9.1
>
Jordan Crouse Dec. 19, 2019, 8:10 p.m. UTC | #2
On Thu, Dec 19, 2019 at 12:58:15PM -0700, Jordan Crouse wrote:
> On Thu, Dec 19, 2019 at 06:44:46PM +0530, Sharat Masetty wrote:


<snip>

> > +
> > +	/*
> > +	 * CNTL1 is used to specify SCID for (CP, TP, VFD, CCU and UBWC
> > +	 * FLAG cache) GPU blocks. This value will be passed along with
> > +	 * the address for any memory transaction from GPU to identify
> > +	 * the sub-cache for that transaction.
> > +	 */
> > +	if (!IS_ERR(llc->gpu_llc_slice)) {
> > +		u32 gpu_scid = llcc_get_slice_id(llc->gpu_llc_slice);
> > +		int i;
> > +
> > +		for (i = 0; i < A6XX_LLC_NUM_GPU_SCIDS; i++)
> > +			llc->cntl1_regval |=
> > +				gpu_scid << (A6XX_GPU_LLC_SCID_NUM_BITS * i);
> 
> As above, i'm not sure a loop is better than just:
> 
> gpu_scid &= 0x1f;
> 
> llc->cntl1_regval = (gpu_scid << 0) || (gpu_scid << 5) | (gpu_scid << 10)
>  | (gpu_scid << 15) | (gpu_scid << 20);
> 
> And I'm not even sure we need do this math here in the first place.

One more question - can you get a valid slice id before activation?

<snip>

Jordan
Sharat Masetty Dec. 20, 2019, 10:10 a.m. UTC | #3
On 2019-12-20 01:28, Jordan Crouse wrote:
> On Thu, Dec 19, 2019 at 06:44:46PM +0530, Sharat Masetty wrote:
>> The last level system cache can be partitioned to 32 different slices
>> of which GPU has two slices preallocated. One slice is used for 
>> caching GPU
>> buffers and the other slice is used for caching the GPU SMMU 
>> pagetables.
>> This patch talks to the core system cache driver to acquire the slice 
>> handles,
>> configure the SCID's to those slices and activates and deactivates the 
>> slices
>> upon GPU power collapse and restore.
>> 
>> Some support from the IOMMU driver is also needed to make use of the
>> system cache. IOMMU_QCOM_SYS_CACHE is a buffer protection flag which 
>> enables
>> caching GPU data buffers in the system cache with memory attributes 
>> such
>> as outer cacheable, read-allocate, write-allocate for buffers. The GPU
>> then has the ability to override a few cacheability parameters which 
>> it
>> does to override write-allocate to write-no-allocate as the GPU 
>> hardware
>> does not benefit much from it.
>> 
>> Similarly DOMAIN_ATTR_QCOM_SYS_CACHE is another domain level attribute
>> used by the IOMMU driver to set the right attributes to cache the 
>> hardware
>> pagetables into the system cache.
>> 
>> Signed-off-by: Sharat Masetty <smasetty@codeaurora.org>
>> ---
>>  drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 122 
>> +++++++++++++++++++++++++++++++++-
>>  drivers/gpu/drm/msm/adreno/a6xx_gpu.h |   9 +++
>>  drivers/gpu/drm/msm/msm_iommu.c       |  13 ++++
>>  drivers/gpu/drm/msm/msm_mmu.h         |   3 +
>>  4 files changed, 146 insertions(+), 1 deletion(-)
>> 
>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c 
>> b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
>> index faff6ff..0c7fdee 100644
>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
>> @@ -9,6 +9,7 @@
>>  #include "a6xx_gmu.xml.h"
>> 
>>  #include <linux/devfreq.h>
>> +#include <linux/soc/qcom/llcc-qcom.h>
>> 
>>  #define GPU_PAS_ID 13
>> 
>> @@ -781,6 +782,117 @@ static void 
>> a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu)
>>  	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
>>  }
>> 
>> +#define A6XX_LLC_NUM_GPU_SCIDS		5
>> +#define A6XX_GPU_LLC_SCID_NUM_BITS	5
> 
> As I mention below, I'm not sure if we need these
> 
>> +#define A6XX_GPU_LLC_SCID_MASK \
>> +	((1 << (A6XX_LLC_NUM_GPU_SCIDS * A6XX_GPU_LLC_SCID_NUM_BITS)) - 1)
>> +
>> +#define A6XX_GPUHTW_LLC_SCID_SHIFT	25
>> +#define A6XX_GPUHTW_LLC_SCID_MASK \
>> +	(((1 << A6XX_GPU_LLC_SCID_NUM_BITS) - 1) << 
>> A6XX_GPUHTW_LLC_SCID_SHIFT)
>> +
> 
> Normally these go into the envytools regmap but if we're going to do 
> these guys
> lets use the power of <linux/bitfield.h> for good.
> 
> #define A6XX_GPU_LLC_SCID GENMASK(24, 0)
> #define A6XX_GPUHTW_LLC_SCID GENMASK(29, 25)
> 
>> +static inline void a6xx_gpu_cx_rmw(struct a6xx_llc *llc,
> 
> Don't mark C functions as inline - let the compiler figure it out for 
> you.
> 
>> +	u32 reg, u32 mask, u32 or)
>> +{
>> +	msm_rmw(llc->mmio + (reg << 2), mask, or);
>> +}
>> +
>> +static void a6xx_llc_deactivate(struct a6xx_llc *llc)
>> +{
>> +	llcc_slice_deactivate(llc->gpu_llc_slice);
>> +	llcc_slice_deactivate(llc->gpuhtw_llc_slice);
>> +}
>> +
>> +static void a6xx_llc_activate(struct a6xx_llc *llc)
>> +{
>> +	if (!llc->mmio)
>> +		return;
>> +
>> +	/* Program the sub-cache ID for all GPU blocks */
>> +	if (!llcc_slice_activate(llc->gpu_llc_slice))
>> +		a6xx_gpu_cx_rmw(llc,
>> +				REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_1,
>> +				A6XX_GPU_LLC_SCID_MASK,
>> +				(llc->cntl1_regval &
>> +				 A6XX_GPU_LLC_SCID_MASK));
> 
> This is out of order with the comments below, but if we store the slice 
> id then
> you could calculate regval here and not have to store it.
> 
>> +
>> +	/* Program the sub-cache ID for the GPU pagetables */
>> +	if (!llcc_slice_activate(llc->gpuhtw_llc_slice))
> 
> val |= FIELD_SET(A6XX_GPUHTW_LLC_SCID, htw_llc_sliceid);
> 
>> +		a6xx_gpu_cx_rmw(llc,
>> +				REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_1,
>> +				A6XX_GPUHTW_LLC_SCID_MASK,
>> +				(llc->cntl1_regval &
>> +				 A6XX_GPUHTW_LLC_SCID_MASK));
> 
> And this could be FIELD_SET(A6XX_GPUHTW_LLC_SCID, sliceid);
> 
> In theory you could just calculate the u32 and write it directly 
> without a rmw.
> In fact, that might be preferable - if the slice activate failed, you 
> don't want
> to run the risk that the scid for htw is still populated.
> 
>> +
>> +	/* Program cacheability overrides */
>> +	a6xx_gpu_cx_rmw(llc, REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_0, 0xF,
>> +		llc->cntl0_regval);
> 
> As below, this could easily be a constant.
> 
>> +}
>> +
>> +static void a6xx_llc_slices_destroy(struct a6xx_llc *llc)
>> +{
>> +	if (llc->mmio)
>> +		iounmap(llc->mmio);
> 
> msm_ioremap returns a devm_ managed resource, so do not use iounmap() 
> to free
> it. Bets to just leave it and let the gpu device handle it when it goes 
> boom.
> 
>> +
>> +	llcc_slice_putd(llc->gpu_llc_slice);
>> +	llcc_slice_putd(llc->gpuhtw_llc_slice);
>> +}
>> +
>> +static int a6xx_llc_slices_init(struct platform_device *pdev,
> T
> This can be void, I don't think we care if it passes or fails.
> 
>> +		struct a6xx_llc *llc)
>> +{
>> +	llc->mmio = msm_ioremap(pdev, "cx_mem", "gpu_cx");
>> +	if (IS_ERR_OR_NULL(llc->mmio))
> 
> msm_ioremap can not return NULL.
> 
>> +		return -ENODEV;
>> +
>> +	llc->gpu_llc_slice = llcc_slice_getd(LLCC_GPU);
>> +	llc->gpuhtw_llc_slice = llcc_slice_getd(LLCC_GPUHTW);
>> +	if (IS_ERR(llc->gpu_llc_slice) && IS_ERR(llc->gpuhtw_llc_slice))
>> +		return -ENODEV;
>> +
>> +	/*
>> +	 * CNTL0 provides options to override the settings for the
>> +	 * read and write allocation policies for the LLC. These
>> +	 * overrides are global for all memory transactions from
>> +	 * the GPU.
>> +	 *
>> +	 * 0x3: read-no-alloc-overridden = 0
>> +	 *      read-no-alloc = 0 - Allocate lines on read miss
>> +	 *      write-no-alloc-overridden = 1
>> +	 *      write-no-alloc = 1 - Do not allocates lines on write miss
>> +	 */
>> +	llc->cntl0_regval = 0x03;
> 
> This is a fixed value isn't it?  We should be able to get away with 
> writing a
> constant.
> 
>> +
>> +	/*
>> +	 * CNTL1 is used to specify SCID for (CP, TP, VFD, CCU and UBWC
>> +	 * FLAG cache) GPU blocks. This value will be passed along with
>> +	 * the address for any memory transaction from GPU to identify
>> +	 * the sub-cache for that transaction.
>> +	 */
>> +	if (!IS_ERR(llc->gpu_llc_slice)) {
>> +		u32 gpu_scid = llcc_get_slice_id(llc->gpu_llc_slice);
>> +		int i;
>> +
>> +		for (i = 0; i < A6XX_LLC_NUM_GPU_SCIDS; i++)
>> +			llc->cntl1_regval |=
>> +				gpu_scid << (A6XX_GPU_LLC_SCID_NUM_BITS * i);
> 
> As above, i'm not sure a loop is better than just:
> 
> gpu_scid &= 0x1f;
> 
> llc->cntl1_regval = (gpu_scid << 0) || (gpu_scid << 5) | (gpu_scid << 
> 10)
>  | (gpu_scid << 15) | (gpu_scid << 20);
> 
> And I'm not even sure we need do this math here in the first place.
> 
>> +	}
>> +
>> +	/*
>> +	 * Set SCID for GPU IOMMU. This will be used to access
>> +	 * page tables that are cached in LLC.
>> +	 */
>> +	if (!IS_ERR(llc->gpuhtw_llc_slice)) {
>> +		u32 gpuhtw_scid = llcc_get_slice_id(llc->gpuhtw_llc_slice);
>> +
>> +		llc->cntl1_regval |=
>> +			gpuhtw_scid << A6XX_GPUHTW_LLC_SCID_SHIFT;
>> +	}
> 
> As above, I think storing the slice id could be more beneficial than 
> calculating
> a value, but if we do calculate a value, use 
> FIELD_SET(A6XX_GPUHTW_LLC_SCID, )
> 
>> +
>> +	return 0;
>> +}
>> +
>>  static int a6xx_pm_resume(struct msm_gpu *gpu)
>>  {
>>  	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
>> @@ -795,6 +907,8 @@ static int a6xx_pm_resume(struct msm_gpu *gpu)
>> 
>>  	msm_gpu_resume_devfreq(gpu);
>> 
>> +	a6xx_llc_activate(&a6xx_gpu->llc);
>> +
>>  	return 0;
>>  }
>> 
>> @@ -803,6 +917,8 @@ static int a6xx_pm_suspend(struct msm_gpu *gpu)
>>  	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
>>  	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
>> 
>> +	a6xx_llc_deactivate(&a6xx_gpu->llc);
>> +
>>  	devfreq_suspend_device(gpu->devfreq.devfreq);
>> 
>>  	/*
>> @@ -851,6 +967,7 @@ static void a6xx_destroy(struct msm_gpu *gpu)
>>  		drm_gem_object_put_unlocked(a6xx_gpu->sqe_bo);
>>  	}
>> 
>> +	a6xx_llc_slices_destroy(&a6xx_gpu->llc);
>>  	a6xx_gmu_remove(a6xx_gpu);
>> 
>>  	adreno_gpu_cleanup(adreno_gpu);
>> @@ -924,7 +1041,10 @@ struct msm_gpu *a6xx_gpu_init(struct drm_device 
>> *dev)
>>  	adreno_gpu->registers = NULL;
>>  	adreno_gpu->reg_offsets = a6xx_register_offsets;
>> 
>> -	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1, 0);
>> +	ret = a6xx_llc_slices_init(pdev, &a6xx_gpu->llc);
>> +
> 
> Confirming we don't care if a6xx_llc_slices_init passes or fails.

Are you suggesting to unconditionally set the memory attributes in 
iommu(see the code below in msm_iommu.c).
We probably wouldn't need this patch too in that case: 
https://patchwork.freedesktop.org/patch/346097/

The return code  is used in the line below to pass 
MMU_FEATURE_USE_SYSTEM_CACHE. Am I missing something here?

> 
>> +	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1,
>> +			ret ? 0 : MMU_FEATURE_USE_SYSTEM_CACHE);
>>  	if (ret) {
>>  		a6xx_destroy(&(a6xx_gpu->base.base));
>>  		return ERR_PTR(ret);
>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h 
>> b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
>> index 7239b8b..09b9ad0 100644
>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
>> @@ -12,6 +12,14 @@
>> 
>>  extern bool hang_debug;
>> 
>> +struct a6xx_llc {
>> +	void __iomem *mmio;
>> +	void *gpu_llc_slice;
>> +	void *gpuhtw_llc_slice;
>> +	u32 cntl0_regval;
> 
> As above, I'm not sure if cntl0 is needed.  Heck, I'm not even sure 
> cntl1 is
> needed - since we could store or query the ids at activate time.
> 
>> +	u32 cntl1_regval;
>> +};
>> +
>>  struct a6xx_gpu {
>>  	struct adreno_gpu base;
>> 
>> @@ -21,6 +29,7 @@ struct a6xx_gpu {
>>  	struct msm_ringbuffer *cur_ring;
>> 
>>  	struct a6xx_gmu gmu;
>> +	struct a6xx_llc llc;
>>  };
>> 
>>  #define to_a6xx_gpu(x) container_of(x, struct a6xx_gpu, base)
>> diff --git a/drivers/gpu/drm/msm/msm_iommu.c 
>> b/drivers/gpu/drm/msm/msm_iommu.c
>> index 8c95c31..4699367 100644
>> --- a/drivers/gpu/drm/msm/msm_iommu.c
>> +++ b/drivers/gpu/drm/msm/msm_iommu.c
>> @@ -27,6 +27,16 @@ static int msm_iommu_attach(struct msm_mmu *mmu, 
>> const char * const *names,
>>  			    int cnt)
>>  {
>>  	struct msm_iommu *iommu = to_msm_iommu(mmu);
>> +	int gpu_htw_llc = 1;
>> +
>> +	/*
>> +	 * This allows GPU to set the bus attributes required
>> +	 * to use system cache on behalf of the iommu page table
>> +	 * walker.
>> +	 */
>> +	if (msm_mmu_has_feature(mmu, MMU_FEATURE_USE_SYSTEM_CACHE))
>> +		iommu_domain_set_attr(iommu->domain,
>> +				DOMAIN_ATTR_QCOM_SYS_CACHE, &gpu_htw_llc);
> 
> We're all okay if this fails?  No harm no foul?
> 
>> 
>>  	return iommu_attach_device(iommu->domain, mmu->dev);
>>  }
>> @@ -45,6 +55,9 @@ static int msm_iommu_map(struct msm_mmu *mmu, 
>> uint64_t iova,
>>  	struct msm_iommu *iommu = to_msm_iommu(mmu);
>>  	size_t ret;
>> 
>> +	if (msm_mmu_has_feature(mmu, MMU_FEATURE_USE_SYSTEM_CACHE))
>> +		prot |= IOMMU_QCOM_SYS_CACHE;
>> +
>>  	ret = iommu_map_sg(iommu->domain, iova, sgt->sgl, sgt->nents, prot);
>>  	WARN_ON(!ret);
>> 
>> diff --git a/drivers/gpu/drm/msm/msm_mmu.h 
>> b/drivers/gpu/drm/msm/msm_mmu.h
>> index 1e4ac36d..3e6bdad 100644
>> --- a/drivers/gpu/drm/msm/msm_mmu.h
>> +++ b/drivers/gpu/drm/msm/msm_mmu.h
>> @@ -18,6 +18,9 @@ struct msm_mmu_funcs {
>>  	void (*destroy)(struct msm_mmu *mmu);
>>  };
>> 
>> +/* MMU features */
>> +#define MMU_FEATURE_USE_SYSTEM_CACHE (1 << 0)
>> +
>>  struct msm_mmu {
>>  	const struct msm_mmu_funcs *funcs;
>>  	struct device *dev;
>> --
>> 1.9.1
>>
Jordan Crouse Dec. 20, 2019, 7:57 p.m. UTC | #4
On Fri, Dec 20, 2019 at 03:40:59PM +0530, smasetty@codeaurora.org wrote:
> On 2019-12-20 01:28, Jordan Crouse wrote:
> >On Thu, Dec 19, 2019 at 06:44:46PM +0530, Sharat Masetty wrote:
> >>The last level system cache can be partitioned to 32 different slices
> >>of which GPU has two slices preallocated. One slice is used for caching
> >>GPU
> >>buffers and the other slice is used for caching the GPU SMMU pagetables.
> >>This patch talks to the core system cache driver to acquire the slice
> >>handles,
> >>configure the SCID's to those slices and activates and deactivates the
> >>slices
> >>upon GPU power collapse and restore.
> >>
> >>Some support from the IOMMU driver is also needed to make use of the
> >>system cache. IOMMU_QCOM_SYS_CACHE is a buffer protection flag which
> >>enables
> >>caching GPU data buffers in the system cache with memory attributes such
> >>as outer cacheable, read-allocate, write-allocate for buffers. The GPU
> >>then has the ability to override a few cacheability parameters which it
> >>does to override write-allocate to write-no-allocate as the GPU hardware
> >>does not benefit much from it.
> >>
> >>Similarly DOMAIN_ATTR_QCOM_SYS_CACHE is another domain level attribute
> >>used by the IOMMU driver to set the right attributes to cache the
> >>hardware
> >>pagetables into the system cache.
> >>
> >>Signed-off-by: Sharat Masetty <smasetty@codeaurora.org>
> >>---
> >> drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 122
> >>+++++++++++++++++++++++++++++++++-
> >> drivers/gpu/drm/msm/adreno/a6xx_gpu.h |   9 +++
> >> drivers/gpu/drm/msm/msm_iommu.c       |  13 ++++
> >> drivers/gpu/drm/msm/msm_mmu.h         |   3 +
> >> 4 files changed, 146 insertions(+), 1 deletion(-)
> >>
> >>diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> >>b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> >>index faff6ff..0c7fdee 100644
> >>--- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> >>+++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> >>@@ -9,6 +9,7 @@
> >> #include "a6xx_gmu.xml.h"
> >>
> >> #include <linux/devfreq.h>
> >>+#include <linux/soc/qcom/llcc-qcom.h>
> >>
> >> #define GPU_PAS_ID 13
> >>
> >>@@ -781,6 +782,117 @@ static void
> >>a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu)
> >> 	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
> >> }
> >>
> >>+#define A6XX_LLC_NUM_GPU_SCIDS		5
> >>+#define A6XX_GPU_LLC_SCID_NUM_BITS	5
> >
> >As I mention below, I'm not sure if we need these
> >
> >>+#define A6XX_GPU_LLC_SCID_MASK \
> >>+	((1 << (A6XX_LLC_NUM_GPU_SCIDS * A6XX_GPU_LLC_SCID_NUM_BITS)) - 1)
> >>+
> >>+#define A6XX_GPUHTW_LLC_SCID_SHIFT	25
> >>+#define A6XX_GPUHTW_LLC_SCID_MASK \
> >>+	(((1 << A6XX_GPU_LLC_SCID_NUM_BITS) - 1) <<
> >>A6XX_GPUHTW_LLC_SCID_SHIFT)
> >>+
> >
> >Normally these go into the envytools regmap but if we're going to do these
> >guys
> >lets use the power of <linux/bitfield.h> for good.
> >
> >#define A6XX_GPU_LLC_SCID GENMASK(24, 0)
> >#define A6XX_GPUHTW_LLC_SCID GENMASK(29, 25)
> >
> >>+static inline void a6xx_gpu_cx_rmw(struct a6xx_llc *llc,
> >
> >Don't mark C functions as inline - let the compiler figure it out for you.
> >
> >>+	u32 reg, u32 mask, u32 or)
> >>+{
> >>+	msm_rmw(llc->mmio + (reg << 2), mask, or);
> >>+}
> >>+
> >>+static void a6xx_llc_deactivate(struct a6xx_llc *llc)
> >>+{
> >>+	llcc_slice_deactivate(llc->gpu_llc_slice);
> >>+	llcc_slice_deactivate(llc->gpuhtw_llc_slice);
> >>+}
> >>+
> >>+static void a6xx_llc_activate(struct a6xx_llc *llc)
> >>+{
> >>+	if (!llc->mmio)
> >>+		return;
> >>+
> >>+	/* Program the sub-cache ID for all GPU blocks */
> >>+	if (!llcc_slice_activate(llc->gpu_llc_slice))
> >>+		a6xx_gpu_cx_rmw(llc,
> >>+				REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_1,
> >>+				A6XX_GPU_LLC_SCID_MASK,
> >>+				(llc->cntl1_regval &
> >>+				 A6XX_GPU_LLC_SCID_MASK));
> >
> >This is out of order with the comments below, but if we store the slice id
> >then
> >you could calculate regval here and not have to store it.
> >
> >>+
> >>+	/* Program the sub-cache ID for the GPU pagetables */
> >>+	if (!llcc_slice_activate(llc->gpuhtw_llc_slice))
> >
> >val |= FIELD_SET(A6XX_GPUHTW_LLC_SCID, htw_llc_sliceid);
> >
> >>+		a6xx_gpu_cx_rmw(llc,
> >>+				REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_1,
> >>+				A6XX_GPUHTW_LLC_SCID_MASK,
> >>+				(llc->cntl1_regval &
> >>+				 A6XX_GPUHTW_LLC_SCID_MASK));
> >
> >And this could be FIELD_SET(A6XX_GPUHTW_LLC_SCID, sliceid);
> >
> >In theory you could just calculate the u32 and write it directly without a
> >rmw.
> >In fact, that might be preferable - if the slice activate failed, you
> >don't want
> >to run the risk that the scid for htw is still populated.
> >
> >>+
> >>+	/* Program cacheability overrides */
> >>+	a6xx_gpu_cx_rmw(llc, REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_0, 0xF,
> >>+		llc->cntl0_regval);
> >
> >As below, this could easily be a constant.
> >
> >>+}
> >>+
> >>+static void a6xx_llc_slices_destroy(struct a6xx_llc *llc)
> >>+{
> >>+	if (llc->mmio)
> >>+		iounmap(llc->mmio);
> >
> >msm_ioremap returns a devm_ managed resource, so do not use iounmap() to
> >free
> >it. Bets to just leave it and let the gpu device handle it when it goes
> >boom.
> >
> >>+
> >>+	llcc_slice_putd(llc->gpu_llc_slice);
> >>+	llcc_slice_putd(llc->gpuhtw_llc_slice);
> >>+}
> >>+
> >>+static int a6xx_llc_slices_init(struct platform_device *pdev,
> >T
> >This can be void, I don't think we care if it passes or fails.
> >
> >>+		struct a6xx_llc *llc)
> >>+{
> >>+	llc->mmio = msm_ioremap(pdev, "cx_mem", "gpu_cx");
> >>+	if (IS_ERR_OR_NULL(llc->mmio))
> >
> >msm_ioremap can not return NULL.
> >
> >>+		return -ENODEV;
> >>+
> >>+	llc->gpu_llc_slice = llcc_slice_getd(LLCC_GPU);
> >>+	llc->gpuhtw_llc_slice = llcc_slice_getd(LLCC_GPUHTW);
> >>+	if (IS_ERR(llc->gpu_llc_slice) && IS_ERR(llc->gpuhtw_llc_slice))
> >>+		return -ENODEV;
> >>+
> >>+	/*
> >>+	 * CNTL0 provides options to override the settings for the
> >>+	 * read and write allocation policies for the LLC. These
> >>+	 * overrides are global for all memory transactions from
> >>+	 * the GPU.
> >>+	 *
> >>+	 * 0x3: read-no-alloc-overridden = 0
> >>+	 *      read-no-alloc = 0 - Allocate lines on read miss
> >>+	 *      write-no-alloc-overridden = 1
> >>+	 *      write-no-alloc = 1 - Do not allocates lines on write miss
> >>+	 */
> >>+	llc->cntl0_regval = 0x03;
> >
> >This is a fixed value isn't it?  We should be able to get away with
> >writing a
> >constant.
> >
> >>+
> >>+	/*
> >>+	 * CNTL1 is used to specify SCID for (CP, TP, VFD, CCU and UBWC
> >>+	 * FLAG cache) GPU blocks. This value will be passed along with
> >>+	 * the address for any memory transaction from GPU to identify
> >>+	 * the sub-cache for that transaction.
> >>+	 */
> >>+	if (!IS_ERR(llc->gpu_llc_slice)) {
> >>+		u32 gpu_scid = llcc_get_slice_id(llc->gpu_llc_slice);
> >>+		int i;
> >>+
> >>+		for (i = 0; i < A6XX_LLC_NUM_GPU_SCIDS; i++)
> >>+			llc->cntl1_regval |=
> >>+				gpu_scid << (A6XX_GPU_LLC_SCID_NUM_BITS * i);
> >
> >As above, i'm not sure a loop is better than just:
> >
> >gpu_scid &= 0x1f;
> >
> >llc->cntl1_regval = (gpu_scid << 0) || (gpu_scid << 5) | (gpu_scid << 10)
> > | (gpu_scid << 15) | (gpu_scid << 20);
> >
> >And I'm not even sure we need do this math here in the first place.
> >
> >>+	}
> >>+
> >>+	/*
> >>+	 * Set SCID for GPU IOMMU. This will be used to access
> >>+	 * page tables that are cached in LLC.
> >>+	 */
> >>+	if (!IS_ERR(llc->gpuhtw_llc_slice)) {
> >>+		u32 gpuhtw_scid = llcc_get_slice_id(llc->gpuhtw_llc_slice);
> >>+
> >>+		llc->cntl1_regval |=
> >>+			gpuhtw_scid << A6XX_GPUHTW_LLC_SCID_SHIFT;
> >>+	}
> >
> >As above, I think storing the slice id could be more beneficial than
> >calculating
> >a value, but if we do calculate a value, use
> >FIELD_SET(A6XX_GPUHTW_LLC_SCID, )
> >
> >>+
> >>+	return 0;
> >>+}
> >>+
> >> static int a6xx_pm_resume(struct msm_gpu *gpu)
> >> {
> >> 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> >>@@ -795,6 +907,8 @@ static int a6xx_pm_resume(struct msm_gpu *gpu)
> >>
> >> 	msm_gpu_resume_devfreq(gpu);
> >>
> >>+	a6xx_llc_activate(&a6xx_gpu->llc);
> >>+
> >> 	return 0;
> >> }
> >>
> >>@@ -803,6 +917,8 @@ static int a6xx_pm_suspend(struct msm_gpu *gpu)
> >> 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> >> 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
> >>
> >>+	a6xx_llc_deactivate(&a6xx_gpu->llc);
> >>+
> >> 	devfreq_suspend_device(gpu->devfreq.devfreq);
> >>
> >> 	/*
> >>@@ -851,6 +967,7 @@ static void a6xx_destroy(struct msm_gpu *gpu)
> >> 		drm_gem_object_put_unlocked(a6xx_gpu->sqe_bo);
> >> 	}
> >>
> >>+	a6xx_llc_slices_destroy(&a6xx_gpu->llc);
> >> 	a6xx_gmu_remove(a6xx_gpu);
> >>
> >> 	adreno_gpu_cleanup(adreno_gpu);
> >>@@ -924,7 +1041,10 @@ struct msm_gpu *a6xx_gpu_init(struct drm_device
> >>*dev)
> >> 	adreno_gpu->registers = NULL;
> >> 	adreno_gpu->reg_offsets = a6xx_register_offsets;
> >>
> >>-	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1, 0);
> >>+	ret = a6xx_llc_slices_init(pdev, &a6xx_gpu->llc);
> >>+
> >
> >Confirming we don't care if a6xx_llc_slices_init passes or fails.
> 
> Are you suggesting to unconditionally set the memory attributes in iommu(see
> the code below in msm_iommu.c).
> We probably wouldn't need this patch too in that case:
> https://patchwork.freedesktop.org/patch/346097/
> 
> The return code  is used in the line below to pass
> MMU_FEATURE_USE_SYSTEM_CACHE. Am I missing something here?

Oh, I see. Please don't do that. Set a separate flag if you need to.

features = 0;

 if (a6xx_llc_slices_init(pdev, &a6xx_gpu->llc))
    features = MMU_FEATURE_USE_SYSTEM_CACHE;

Hiding ret in a function that also sets ret has a tendency to confuse people
like me.

Jordan
diff mbox series

Patch

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
index faff6ff..0c7fdee 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
@@ -9,6 +9,7 @@ 
 #include "a6xx_gmu.xml.h"

 #include <linux/devfreq.h>
+#include <linux/soc/qcom/llcc-qcom.h>

 #define GPU_PAS_ID 13

@@ -781,6 +782,117 @@  static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu)
 	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
 }

+#define A6XX_LLC_NUM_GPU_SCIDS		5
+#define A6XX_GPU_LLC_SCID_NUM_BITS	5
+
+#define A6XX_GPU_LLC_SCID_MASK \
+	((1 << (A6XX_LLC_NUM_GPU_SCIDS * A6XX_GPU_LLC_SCID_NUM_BITS)) - 1)
+
+#define A6XX_GPUHTW_LLC_SCID_SHIFT	25
+#define A6XX_GPUHTW_LLC_SCID_MASK \
+	(((1 << A6XX_GPU_LLC_SCID_NUM_BITS) - 1) << A6XX_GPUHTW_LLC_SCID_SHIFT)
+
+static inline void a6xx_gpu_cx_rmw(struct a6xx_llc *llc,
+	u32 reg, u32 mask, u32 or)
+{
+	msm_rmw(llc->mmio + (reg << 2), mask, or);
+}
+
+static void a6xx_llc_deactivate(struct a6xx_llc *llc)
+{
+	llcc_slice_deactivate(llc->gpu_llc_slice);
+	llcc_slice_deactivate(llc->gpuhtw_llc_slice);
+}
+
+static void a6xx_llc_activate(struct a6xx_llc *llc)
+{
+	if (!llc->mmio)
+		return;
+
+	/* Program the sub-cache ID for all GPU blocks */
+	if (!llcc_slice_activate(llc->gpu_llc_slice))
+		a6xx_gpu_cx_rmw(llc,
+				REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_1,
+				A6XX_GPU_LLC_SCID_MASK,
+				(llc->cntl1_regval &
+				 A6XX_GPU_LLC_SCID_MASK));
+
+	/* Program the sub-cache ID for the GPU pagetables */
+	if (!llcc_slice_activate(llc->gpuhtw_llc_slice))
+		a6xx_gpu_cx_rmw(llc,
+				REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_1,
+				A6XX_GPUHTW_LLC_SCID_MASK,
+				(llc->cntl1_regval &
+				 A6XX_GPUHTW_LLC_SCID_MASK));
+
+	/* Program cacheability overrides */
+	a6xx_gpu_cx_rmw(llc, REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_0, 0xF,
+		llc->cntl0_regval);
+}
+
+static void a6xx_llc_slices_destroy(struct a6xx_llc *llc)
+{
+	if (llc->mmio)
+		iounmap(llc->mmio);
+
+	llcc_slice_putd(llc->gpu_llc_slice);
+	llcc_slice_putd(llc->gpuhtw_llc_slice);
+}
+
+static int a6xx_llc_slices_init(struct platform_device *pdev,
+		struct a6xx_llc *llc)
+{
+	llc->mmio = msm_ioremap(pdev, "cx_mem", "gpu_cx");
+	if (IS_ERR_OR_NULL(llc->mmio))
+		return -ENODEV;
+
+	llc->gpu_llc_slice = llcc_slice_getd(LLCC_GPU);
+	llc->gpuhtw_llc_slice = llcc_slice_getd(LLCC_GPUHTW);
+	if (IS_ERR(llc->gpu_llc_slice) && IS_ERR(llc->gpuhtw_llc_slice))
+		return -ENODEV;
+
+	/*
+	 * CNTL0 provides options to override the settings for the
+	 * read and write allocation policies for the LLC. These
+	 * overrides are global for all memory transactions from
+	 * the GPU.
+	 *
+	 * 0x3: read-no-alloc-overridden = 0
+	 *      read-no-alloc = 0 - Allocate lines on read miss
+	 *      write-no-alloc-overridden = 1
+	 *      write-no-alloc = 1 - Do not allocates lines on write miss
+	 */
+	llc->cntl0_regval = 0x03;
+
+	/*
+	 * CNTL1 is used to specify SCID for (CP, TP, VFD, CCU and UBWC
+	 * FLAG cache) GPU blocks. This value will be passed along with
+	 * the address for any memory transaction from GPU to identify
+	 * the sub-cache for that transaction.
+	 */
+	if (!IS_ERR(llc->gpu_llc_slice)) {
+		u32 gpu_scid = llcc_get_slice_id(llc->gpu_llc_slice);
+		int i;
+
+		for (i = 0; i < A6XX_LLC_NUM_GPU_SCIDS; i++)
+			llc->cntl1_regval |=
+				gpu_scid << (A6XX_GPU_LLC_SCID_NUM_BITS * i);
+	}
+
+	/*
+	 * Set SCID for GPU IOMMU. This will be used to access
+	 * page tables that are cached in LLC.
+	 */
+	if (!IS_ERR(llc->gpuhtw_llc_slice)) {
+		u32 gpuhtw_scid = llcc_get_slice_id(llc->gpuhtw_llc_slice);
+
+		llc->cntl1_regval |=
+			gpuhtw_scid << A6XX_GPUHTW_LLC_SCID_SHIFT;
+	}
+
+	return 0;
+}
+
 static int a6xx_pm_resume(struct msm_gpu *gpu)
 {
 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
@@ -795,6 +907,8 @@  static int a6xx_pm_resume(struct msm_gpu *gpu)

 	msm_gpu_resume_devfreq(gpu);

+	a6xx_llc_activate(&a6xx_gpu->llc);
+
 	return 0;
 }

@@ -803,6 +917,8 @@  static int a6xx_pm_suspend(struct msm_gpu *gpu)
 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);

+	a6xx_llc_deactivate(&a6xx_gpu->llc);
+
 	devfreq_suspend_device(gpu->devfreq.devfreq);

 	/*
@@ -851,6 +967,7 @@  static void a6xx_destroy(struct msm_gpu *gpu)
 		drm_gem_object_put_unlocked(a6xx_gpu->sqe_bo);
 	}

+	a6xx_llc_slices_destroy(&a6xx_gpu->llc);
 	a6xx_gmu_remove(a6xx_gpu);

 	adreno_gpu_cleanup(adreno_gpu);
@@ -924,7 +1041,10 @@  struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
 	adreno_gpu->registers = NULL;
 	adreno_gpu->reg_offsets = a6xx_register_offsets;

-	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1, 0);
+	ret = a6xx_llc_slices_init(pdev, &a6xx_gpu->llc);
+
+	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1,
+			ret ? 0 : MMU_FEATURE_USE_SYSTEM_CACHE);
 	if (ret) {
 		a6xx_destroy(&(a6xx_gpu->base.base));
 		return ERR_PTR(ret);
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
index 7239b8b..09b9ad0 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
@@ -12,6 +12,14 @@ 

 extern bool hang_debug;

+struct a6xx_llc {
+	void __iomem *mmio;
+	void *gpu_llc_slice;
+	void *gpuhtw_llc_slice;
+	u32 cntl0_regval;
+	u32 cntl1_regval;
+};
+
 struct a6xx_gpu {
 	struct adreno_gpu base;

@@ -21,6 +29,7 @@  struct a6xx_gpu {
 	struct msm_ringbuffer *cur_ring;

 	struct a6xx_gmu gmu;
+	struct a6xx_llc llc;
 };

 #define to_a6xx_gpu(x) container_of(x, struct a6xx_gpu, base)
diff --git a/drivers/gpu/drm/msm/msm_iommu.c b/drivers/gpu/drm/msm/msm_iommu.c
index 8c95c31..4699367 100644
--- a/drivers/gpu/drm/msm/msm_iommu.c
+++ b/drivers/gpu/drm/msm/msm_iommu.c
@@ -27,6 +27,16 @@  static int msm_iommu_attach(struct msm_mmu *mmu, const char * const *names,
 			    int cnt)
 {
 	struct msm_iommu *iommu = to_msm_iommu(mmu);
+	int gpu_htw_llc = 1;
+
+	/*
+	 * This allows GPU to set the bus attributes required
+	 * to use system cache on behalf of the iommu page table
+	 * walker.
+	 */
+	if (msm_mmu_has_feature(mmu, MMU_FEATURE_USE_SYSTEM_CACHE))
+		iommu_domain_set_attr(iommu->domain,
+				DOMAIN_ATTR_QCOM_SYS_CACHE, &gpu_htw_llc);

 	return iommu_attach_device(iommu->domain, mmu->dev);
 }
@@ -45,6 +55,9 @@  static int msm_iommu_map(struct msm_mmu *mmu, uint64_t iova,
 	struct msm_iommu *iommu = to_msm_iommu(mmu);
 	size_t ret;

+	if (msm_mmu_has_feature(mmu, MMU_FEATURE_USE_SYSTEM_CACHE))
+		prot |= IOMMU_QCOM_SYS_CACHE;
+
 	ret = iommu_map_sg(iommu->domain, iova, sgt->sgl, sgt->nents, prot);
 	WARN_ON(!ret);

diff --git a/drivers/gpu/drm/msm/msm_mmu.h b/drivers/gpu/drm/msm/msm_mmu.h
index 1e4ac36d..3e6bdad 100644
--- a/drivers/gpu/drm/msm/msm_mmu.h
+++ b/drivers/gpu/drm/msm/msm_mmu.h
@@ -18,6 +18,9 @@  struct msm_mmu_funcs {
 	void (*destroy)(struct msm_mmu *mmu);
 };

+/* MMU features */
+#define MMU_FEATURE_USE_SYSTEM_CACHE (1 << 0)
+
 struct msm_mmu {
 	const struct msm_mmu_funcs *funcs;
 	struct device *dev;