drm/msm/a6xx: System Cache Support
Subsystems (5)
cover drm/msm/a6xx: System Cache Support
Some hardware variants contain a system level cache or the last level cache(llc). This cache is typically a large block which is shared by multiple clients on the SOC. GPU uses the system cache to cache both the GPU data buffers(like textures) as well the SMMU pagetables. This helps with improved render performance as well as lower power consumption by reducing the bus traffic to the system memory.
The system cache architecture allows the cache to be split into slices which then be used by multiple SOC clients. This patch series is an effort to enable and use two of those slices perallocated for the GPU, one for the GPU data buffers and another for the GPU SMMU hardware pagetables.
To enable the system cache driver, add [1] to your stack if not already present. Please review.
[1] https://lore.kernel.org/patchwork/patch/1165298/
Jordan Crouse (1):
iommu/arm-smmu: Pass io_pgtable_cfg to impl specific init_context
Sharat Masetty (3):
drm/msm: rearrange the gpu_rmw() function drm/msm: Pass mmu features to generic layers drm/msm/a6xx: Add support for using system cache(LLC)
Vivek Gautam (1):
iommu/arm-smmu: Add domain attribute for QCOM system cache
drivers/gpu/drm/msm/adreno/a2xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a3xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a4xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 122 +++++++++++++++++++++++++++++++- drivers/gpu/drm/msm/adreno/a6xx_gpu.h | 9 +++ drivers/gpu/drm/msm/adreno/adreno_gpu.c | 4 +- drivers/gpu/drm/msm/adreno/adreno_gpu.h | 2 +- drivers/gpu/drm/msm/msm_drv.c | 8 +++ drivers/gpu/drm/msm/msm_drv.h | 1 + drivers/gpu/drm/msm/msm_gpu.c | 6 +- drivers/gpu/drm/msm/msm_gpu.h | 6 +- drivers/gpu/drm/msm/msm_iommu.c | 13 ++++ drivers/gpu/drm/msm/msm_mmu.h | 14 ++++ drivers/iommu/arm-smmu-impl.c | 3 +- drivers/iommu/arm-smmu-qcom.c | 10 +++ drivers/iommu/arm-smmu.c | 25 +++++-- drivers/iommu/arm-smmu.h | 4 +- include/linux/iommu.h | 1 + 19 files changed, 216 insertions(+), 20 deletions(-)
--
1.9.1
PATCH 1/5 iommu/arm-smmu: Pass io_pgtable_cfg to impl specific init_context
From: Jordan Crouse <[email protected]>
Pass the propposed io_pgtable_cfg to the implementation specific init_context() function to give the implementation an opportunity to to modify it before it gets passed to io-pgtable.
--- drivers/iommu/arm-smmu-impl.c | 3 ++- drivers/iommu/arm-smmu.c | 11 ++++++----- drivers/iommu/arm-smmu.h | 3 ++- 3 files changed, 10 insertions(+), 7 deletions(-)
drivers/iommu/arm-smmu-impl.c +2 -1
@@ -68,7 +68,8 @@ static int cavium_cfg_probe(struct arm_smmu_device *smmu)return 0 ; } -static int cavium_init_context (struct arm_smmu_domain *smmu_domain) +static int cavium_init_context (struct arm_smmu_domain *smmu_domain, +struct io_pgtable_cfg *pgtbl_cfg) {struct cavium_smmu *cs =container_of (smmu_domain->smmu, struct cavium_smmu, smmu);
drivers/iommu/arm-smmu.c +6 -5
@@ -758,11 +758,6 @@ static int arm_smmu_init_domain_context(struct iommu_domain *domain, cfg->asid = cfg->cbndx; smmu_domain->smmu = smmu; -if (smmu->impl && smmu->impl->init_context) { - ret = smmu->impl->init_context (smmu_domain); -if (ret) - goto out_unlock; - } smmu_domain->pgtbl_cfg = (struct io_pgtable_cfg ) { .pgsize_bitmap = smmu->pgsize_bitmap,
@@ -773,6 +768,12 @@ static int arm_smmu_init_domain_context(struct iommu_domain *domain, .iommu_dev = smmu->dev, }; +if (smmu->impl && smmu->impl->init_context) { + ret = smmu->impl->init_context (smmu_domain, &smmu_domain->pgtbl_cfg); +if (ret) + goto out_unlock; + } +if (smmu_domain->non_strict) smmu_domain->pgtbl_cfg.quirks |=IO_PGTABLE_QUIRK_NON_STRICT ;
drivers/iommu/arm-smmu.h +2 -1
@@ -335,7 +335,8 @@ struct arm_smmu_impl { u64 val);int (*cfg_probe)(struct arm_smmu_device *smmu);int (*reset)(struct arm_smmu_device *smmu); -int (*init_context)(struct arm_smmu_domain *smmu_domain); +int (*init_context)(struct arm_smmu_domain *smmu_domain, +struct io_pgtable_cfg *pgtbl_cfg); };static inline void __iomem *arm_smmu_page (struct arm_smmu_device *smmu,int n)
--
1.9.1
PATCH 2/5 iommu/arm-smmu: Add domain attribute for QCOM system cache
In reply to Sharat MasettyFrom: Vivek Gautam <[email protected]>
Add iommu domain attribute for using system cache aka last level cache on QCOM SoCs by client drivers like GPU to set right attributes for caching the hardware pagetables into the system cache.
--- drivers/iommu/arm-smmu-qcom.c | 10 ++++++++++ drivers/iommu/arm-smmu.c | 14 ++++++++++++++ drivers/iommu/arm-smmu.h | 1 + include/linux/iommu.h | 1 + 4 files changed, 26 insertions(+)
drivers/iommu/arm-smmu-qcom.c +10 -0
@@ -30,7 +30,17 @@ static int qcom_sdm845_smmu500_reset(struct arm_smmu_device *smmu)return ret; } +static int qcom_smmu_init_context (struct arm_smmu_domain *smmu_domain, +struct io_pgtable_cfg *pgtbl_cfg) +{ +if (smmu_domain->sys_cache) + pgtbl_cfg->coherent_walk = false; + +return 0 ; +} +static const struct arm_smmu_impl qcom_smmu_impl = { + .init_context = qcom_smmu_init_context, .reset = qcom_sdm845_smmu500_reset, };
drivers/iommu/arm-smmu.c +14 -0
@@ -1466,6 +1466,9 @@ static int arm_smmu_domain_get_attr(struct iommu_domain *domain,case DOMAIN_ATTR_NESTING : *(int *)data = (smmu_domain->stage ==ARM_SMMU_DOMAIN_NESTED );return 0 ; +case DOMAIN_ATTR_QCOM_SYS_CACHE : + *((int *)data) = smmu_domain->sys_cache; +return 0 ;default :return -ENODEV ; }
@@ -1506,6 +1509,17 @@ static int arm_smmu_domain_set_attr(struct iommu_domain *domain,else smmu_domain->stage =ARM_SMMU_DOMAIN_S1 ;break ; +case DOMAIN_ATTR_QCOM_SYS_CACHE : +if (smmu_domain->smmu) { + ret = -EPERM ; + goto out_unlock; + } + +if (*((int *)data)) + smmu_domain->sys_cache = true; +else + smmu_domain->sys_cache = false; +break ;default : ret = -ENODEV ; }
drivers/iommu/arm-smmu.h +1 -0
@@ -322,6 +322,7 @@ struct arm_smmu_domain {struct mutex init_mutex;/* Protects smmu pointer */ spinlock_t cb_lock;/* Serialises ATS1* ops and TLB syncs */ struct iommu_domain domain; +bool sys_cache; };
include/linux/iommu.h +1 -0
@@ -127,6 +127,7 @@ enum iommu_attr {DOMAIN_ATTR_FSL_PAMUV1 ,DOMAIN_ATTR_NESTING ,/* two stages of translation */ DOMAIN_ATTR_DMA_USE_FLUSH_QUEUE , +DOMAIN_ATTR_QCOM_SYS_CACHE , DOMAIN_ATTR_MAX, };
Signature · 1 line
1.9.1
PATCH 3/5 drm/msm: rearrange the gpu_rmw() function
In reply to Sharat MasettyThe register read-modify-write construct is generic enough that it can be used by other subsystems as needed, create a more generic rmw() function and have the gpu_rmw() use this new function.
--- drivers/gpu/drm/msm/msm_drv.c | 8 ++++++++ drivers/gpu/drm/msm/msm_drv.h | 1 + drivers/gpu/drm/msm/msm_gpu.h | 5 +---- 3 files changed, 10 insertions(+), 4 deletions(-)
drivers/gpu/drm/msm/msm_drv.c +8 -0
@@ -165,6 +165,14 @@ u32 msm_readl(const void __iomem *addr)return val; } +void msm_rmw (void __iomem *addr,u32 mask,u32 or) +{ +u32 val =msm_readl (addr); + + val &= ~mask; +msm_writel (val | or, addr); +} +struct msm_vblank_work {struct work_struct work;int crtc_id;
drivers/gpu/drm/msm/msm_drv.h +1 -0
@@ -409,6 +409,7 @@ void __iomem *msm_ioremap(struct platform_device *pdev, const char *name, const char *dbgname);void msm_writel (u32 data,void __iomem *addr);u32 msm_readl (const void __iomem *addr); +void msm_rmw (void __iomem *addr,u32 mask,u32 or);struct msm_gpu_submitqueue ;int msm_submitqueue_init (struct drm_device *drm,struct msm_file_private *ctx);
drivers/gpu/drm/msm/msm_gpu.h +1 -4
@@ -223,10 +223,7 @@ static inline u32 gpu_read(struct msm_gpu *gpu, u32 reg)static inline void gpu_rmw (struct msm_gpu *gpu,u32 reg,u32 mask,u32 or) { -uint32_t val =gpu_read (gpu, reg); - - val &= ~mask; -gpu_write (gpu, reg, val | or); +msm_rmw (gpu->mmio + (reg <<2 ), mask, or); }static inline u64 gpu_read64 (struct msm_gpu *gpu,u32 lo,u32 hi)
Signature · 1 line
1.9.1
PATCH 4/5 drm/msm: Pass mmu features to generic layers
In reply to Sharat MasettyAllow different Adreno targets the ability to pass specific mmu features to the generic layers. This will help conditionally configure certain iommu features for certain Adreno targets.
Also Add a few simple support functions to support a bitmask of features that a specific MMU implementation supports.
--- drivers/gpu/drm/msm/adreno/a2xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a3xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a4xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/adreno_gpu.c | 4 +++- drivers/gpu/drm/msm/adreno/adreno_gpu.h | 2 +- drivers/gpu/drm/msm/msm_gpu.c | 6 ++++-- drivers/gpu/drm/msm/msm_gpu.h | 1 + drivers/gpu/drm/msm/msm_mmu.h | 11 +++++++++++ 10 files changed, 25 insertions(+), 9 deletions(-)
drivers/gpu/drm/msm/adreno/a2xx_gpu.c +1 -1
@@ -472,7 +472,7 @@ struct msm_gpu *a2xx_gpu_init(struct drm_device *dev) adreno_gpu->reg_offsets = a2xx_register_offsets; - ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ); + ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ,0 );if (ret) goto fail;
drivers/gpu/drm/msm/adreno/a3xx_gpu.c +1 -1
@@ -488,7 +488,7 @@ struct msm_gpu *a3xx_gpu_init(struct drm_device *dev) adreno_gpu->registers = a3xx_registers; adreno_gpu->reg_offsets = a3xx_register_offsets; - ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ); + ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ,0 );if (ret) goto fail;
drivers/gpu/drm/msm/adreno/a4xx_gpu.c +1 -1
@@ -572,7 +572,7 @@ struct msm_gpu *a4xx_gpu_init(struct drm_device *dev) adreno_gpu->registers = a4xx_registers; adreno_gpu->reg_offsets = a4xx_register_offsets; - ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ); + ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ,0 );if (ret) goto fail;
drivers/gpu/drm/msm/adreno/a5xx_gpu.c +1 -1
@@ -1445,7 +1445,7 @@ struct msm_gpu *a5xx_gpu_init(struct drm_device *dev)check_speed_bin (&pdev->dev); - ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,4 ); + ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,4 ,0 );if (ret) {a5xx_destroy (&(a5xx_gpu->base.base));return ERR_PTR (ret);
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +1 -1
@@ -924,7 +924,7 @@ 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 ); + ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ,0 );if (ret) {a6xx_destroy (&(a6xx_gpu->base.base));return ERR_PTR (ret);
drivers/gpu/drm/msm/adreno/adreno_gpu.c +3 -1
@@ -895,7 +895,8 @@ static int adreno_get_pwrlevels(struct device *dev,int adreno_gpu_init (struct drm_device *drm,struct platform_device *pdev,struct adreno_gpu *adreno_gpu, -const struct adreno_gpu_funcs *funcs,int nr_rings) +const struct adreno_gpu_funcs *funcs,int nr_rings, +u32 mmu_features) {struct adreno_platform_config *config = pdev->dev.platform_data;struct msm_gpu_config adreno_gpu_config = {0 };
@@ -916,6 +917,7 @@ int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev, adreno_gpu_config.va_end =SZ_16M +0xfff *SZ_64K ; adreno_gpu_config.nr_rings = nr_rings; + adreno_gpu_config.mmu_features = mmu_features;adreno_get_pwrlevels (&pdev->dev, gpu);
drivers/gpu/drm/msm/adreno/adreno_gpu.h +1 -1
@@ -248,7 +248,7 @@ void adreno_show(struct msm_gpu *gpu, struct msm_gpu_state *state,int adreno_gpu_init (struct drm_device *drm,struct platform_device *pdev,struct adreno_gpu *gpu,const struct adreno_gpu_funcs *funcs, -int nr_rings); +int nr_rings,u32 mmu_features);void adreno_gpu_cleanup (struct adreno_gpu *gpu);int adreno_load_fw (struct adreno_gpu *adreno_gpu);
drivers/gpu/drm/msm/msm_gpu.c +4 -2
@@ -804,7 +804,7 @@ static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)static struct msm_gem_address_space *msm_gpu_create_address_space (struct msm_gpu *gpu,struct platform_device *pdev, -uint64_t va_start,uint64_t va_end) +uint64_t va_start,uint64_t va_end,u32 mmu_features) {struct msm_gem_address_space *aspace;int ret;
@@ -838,6 +838,8 @@ static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)return ERR_CAST (aspace); } +msm_mmu_set_feature (aspace->mmu, mmu_features); + ret = aspace->mmu->funcs->attach (aspace->mmu,NULL ,0 );if (ret) {msm_gem_address_space_put (aspace);
@@ -920,7 +922,7 @@ int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,msm_devfreq_init (gpu); gpu->aspace =msm_gpu_create_address_space (gpu, pdev, - config->va_start, config->va_end); + config->va_start, config->va_end, config->mmu_features);if (gpu->aspace ==NULL )DRM_DEV_INFO (drm->dev,"%s: no IOMMU, fallback to VRAM carveout!\n" , name);
drivers/gpu/drm/msm/msm_gpu.h +1 -0
@@ -24,6 +24,7 @@ struct msm_gpu_config {uint64_t va_start;uint64_t va_end;unsigned int nr_rings; +u32 mmu_features; };/* So far, with hardware that I've seen to date, we can have:
drivers/gpu/drm/msm/msm_mmu.h +11 -0
@@ -23,6 +23,7 @@ struct msm_mmu {struct device *dev;int (*handler)(void *arg,unsigned long iova,int flags);void *arg; +u32 features; }; static inline void msm_mmu_init(struct msm_mmu *mmu, struct device *dev,
@@ -45,4 +46,14 @@ static inline void msm_mmu_set_fault_handler(struct msm_mmu *mmu, void *arg,void msm_gpummu_params (struct msm_mmu *mmu,dma_addr_t *pt_base,dma_addr_t *tran_error); +static inline void msm_mmu_set_feature (struct msm_mmu *mmu,u32 feature) +{ + mmu->features |= feature; +} + +static inline bool msm_mmu_has_feature (struct msm_mmu *mmu,u32 feature) +{ +return (mmu->features & feature) ? true : false; +} +#endif /* __MSM_MMU_H__ */
Signature · 1 line
1.9.1
Re: [PATCH 4/5] drm/msm: Pass mmu features to generic layers
Quoted 7 lines · Sharat Masetty
Allow different Adreno targets the ability to pass specific mmu features to the generic layers. This will help conditionally configure certain iommu features for certain Adreno targets.
Also Add a few simple support functions to support a bitmask of features that a specific MMU implementation supports.
This whole change could benefit from [1] which makes the address space creation target specific.
That would get rid of most of the blobs. Further more, if you took part of [2] that set up the mmu inside of the target specific code (skipping over the SPLIT_PAGETABLE stuff for now) you could set mmu->features directly and not need a helper function to do it.
[1] https://patchwork.freedesktop.org/patch/342170/
[2] https://patchwork.freedesktop.org/patch/342173/
Jordan
Quoted 189 lines
Signed-off-by Sharat Masetty <[email protected]>--- drivers/gpu/drm/msm/adreno/a2xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a3xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a4xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 2 +- drivers/gpu/drm/msm/adreno/adreno_gpu.c | 4 +++- drivers/gpu/drm/msm/adreno/adreno_gpu.h | 2 +- drivers/gpu/drm/msm/msm_gpu.c | 6 ++++-- drivers/gpu/drm/msm/msm_gpu.h | 1 + drivers/gpu/drm/msm/msm_mmu.h | 11 +++++++++++ 10 files changed, 25 insertions(+), 9 deletions(-)drivers/gpu/drm/msm/adreno/a2xx_gpu.c +1
-1@@ -472,7 +472,7 @@ struct msm_gpu *a2xx_gpu_init(struct drm_device *dev) adreno_gpu->reg_offsets = a2xx_register_offsets; - ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ); + ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ,0 );if (ret) goto fail;drivers/gpu/drm/msm/adreno/a3xx_gpu.c +1
-1@@ -488,7 +488,7 @@ struct msm_gpu *a3xx_gpu_init(struct drm_device *dev) adreno_gpu->registers = a3xx_registers; adreno_gpu->reg_offsets = a3xx_register_offsets; - ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ); + ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ,0 );if (ret) goto fail;drivers/gpu/drm/msm/adreno/a4xx_gpu.c +1
-1@@ -572,7 +572,7 @@ struct msm_gpu *a4xx_gpu_init(struct drm_device *dev) adreno_gpu->registers = a4xx_registers; adreno_gpu->reg_offsets = a4xx_register_offsets; - ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ); + ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ,0 );if (ret) goto fail;drivers/gpu/drm/msm/adreno/a5xx_gpu.c +1
-1@@ -1445,7 +1445,7 @@ struct msm_gpu *a5xx_gpu_init(struct drm_device *dev)check_speed_bin (&pdev->dev); - ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,4 ); + ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,4 ,0 );if (ret) {a5xx_destroy (&(a5xx_gpu->base.base));return ERR_PTR (ret);drivers/gpu/drm/msm/adreno/a6xx_gpu.c +1
-1@@ -924,7 +924,7 @@ 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 ); + ret =adreno_gpu_init (dev, pdev, adreno_gpu, &funcs,1 ,0 );if (ret) {a6xx_destroy (&(a6xx_gpu->base.base));return ERR_PTR (ret);drivers/gpu/drm/msm/adreno/adreno_gpu.c +3
-1@@ -895,7 +895,8 @@ static int adreno_get_pwrlevels(struct device *dev,int adreno_gpu_init (struct drm_device *drm,struct platform_device *pdev,struct adreno_gpu *adreno_gpu, -const struct adreno_gpu_funcs *funcs,int nr_rings) +const struct adreno_gpu_funcs *funcs,int nr_rings, +u32 mmu_features) {struct adreno_platform_config *config = pdev->dev.platform_data;struct msm_gpu_config adreno_gpu_config = {0 };@@ -916,6 +917,7 @@ int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev, adreno_gpu_config.va_end =SZ_16M +0xfff *SZ_64K ; adreno_gpu_config.nr_rings = nr_rings; + adreno_gpu_config.mmu_features = mmu_features;adreno_get_pwrlevels (&pdev->dev, gpu);drivers/gpu/drm/msm/adreno/adreno_gpu.h +1
-1@@ -248,7 +248,7 @@ void adreno_show(struct msm_gpu *gpu, struct msm_gpu_state *state,int adreno_gpu_init (struct drm_device *drm,struct platform_device *pdev,struct adreno_gpu *gpu,const struct adreno_gpu_funcs *funcs, -int nr_rings); +int nr_rings,u32 mmu_features);void adreno_gpu_cleanup (struct adreno_gpu *gpu);int adreno_load_fw (struct adreno_gpu *adreno_gpu);drivers/gpu/drm/msm/msm_gpu.c +4
-2@@ -804,7 +804,7 @@ static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)static struct msm_gem_address_space *msm_gpu_create_address_space (struct msm_gpu *gpu,struct platform_device *pdev, -uint64_t va_start,uint64_t va_end) +uint64_t va_start,uint64_t va_end,u32 mmu_features) {struct msm_gem_address_space *aspace;int ret;@@ -838,6 +838,8 @@ static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)return ERR_CAST (aspace); } +msm_mmu_set_feature (aspace->mmu, mmu_features); + ret = aspace->mmu->funcs->attach (aspace->mmu,NULL ,0 );if (ret) {msm_gem_address_space_put (aspace);@@ -920,7 +922,7 @@ int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,msm_devfreq_init (gpu); gpu->aspace =msm_gpu_create_address_space (gpu, pdev, - config->va_start, config->va_end); + config->va_start, config->va_end, config->mmu_features);drivers/gpu/drm/msm/msm_gpu.c +0
-0if (gpu->aspace ==NULL )DRM_DEV_INFO (drm->dev,"%s: no IOMMU, fallback to VRAM carveout!\n" , name);drivers/gpu/drm/msm/msm_gpu.h +1
-0@@ -24,6 +24,7 @@ struct msm_gpu_config {uint64_t va_start;uint64_t va_end;unsigned int nr_rings; +u32 mmu_features; };/* So far, with hardware that I've seen to date, we can have: drivers/gpu/drm/msm/msm_mmu.h +11
-0@@ -23,6 +23,7 @@ struct msm_mmu {struct device *dev;int (*handler)(void *arg,unsigned long iova,int flags);void *arg; +u32 features; }; static inline void msm_mmu_init(struct msm_mmu *mmu, struct device *dev,@@ -45,4 +46,14 @@ static inline void msm_mmu_set_fault_handler(struct msm_mmu *mmu, void *arg,void msm_gpummu_params (struct msm_mmu *mmu,dma_addr_t *pt_base,dma_addr_t *tran_error); +static inline void msm_mmu_set_feature (struct msm_mmu *mmu,u32 feature) +{ + mmu->features |= feature; +} + +static inline bool msm_mmu_has_feature (struct msm_mmu *mmu,u32 feature) +{ +return (mmu->features & feature) ? true : false; +} +#endif /* __MSM_MMU_H__ */ Signature · 1 line
1.9.1
Signature · 6 lines
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project _______________________________________________ dri-devel mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/dri-devel
PATCH 5/5 drm/msm/a6xx: Add support for using system cache(LLC)
In reply to Sharat MasettyThe 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.
--- 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(-)
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +121 -1
@@ -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);
drivers/gpu/drm/msm/adreno/a6xx_gpu.h +9 -0
@@ -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)
drivers/gpu/drm/msm/msm_iommu.c +13 -0
@@ -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);
drivers/gpu/drm/msm/msm_mmu.h +3 -0
@@ -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
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [PATCH 5/5] drm/msm/a6xx: Add support for using system cache(LLC)
Quoted 40 lines · Sharat Masetty
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_CACHEis 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_CACHEis 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 <[email protected]>--- 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(-)drivers/gpu/drm/msm/adreno/a6xx_gpu.c +1
-0@@ -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
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 786
Quoted 5 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +2
-0@@ -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
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 794
Quoted 7 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +7
-0+#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)
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 795
Quoted 1 line
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +1
-0+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.
Quoted 13 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +13
-0+ 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) +{
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 818
Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +10
-0+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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 821
Quoted 3 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +3
-0+ +/* 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);
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 826
Quoted 5 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +5
-0+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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 830
Quoted 4 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4
-0+ +/* 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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 836
Quoted 6 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +6
-0+} + +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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 842
Quoted 6 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +6
-0+ +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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 846
Quoted 4 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4
-0+ 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.
Quoted 9 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +9
-0+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
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 865
Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +10
-0+* 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.
Quoted 4 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4
-0+ +/* +* CNTL1 is used to specify SCID for (CP, TP, VFD, CCU and UBWC +* FLAG cache) GPU blocks. This value will be passed along with
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 879
Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +10
-0+* 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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 891
Quoted 12 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +12
-0+ } + +/* +* 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, )
Quoted 34 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +9
-0+ +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);
drivers/gpu/drm/msm/adreno/a6xx_gpu.c struct msm_gpu *a6xx_gpu_init(struct drm_device *dev) · line 1045
Quoted 6 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +2
-1@@ -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.
Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +2
-0+ 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);drivers/gpu/drm/msm/adreno/a6xx_gpu.h +0
-0
drivers/gpu/drm/msm/adreno/a6xx_gpu.h line 19
Quoted 8 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.h +5
-0@@ -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.
Quoted 22 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.h +4
-0+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)drivers/gpu/drm/msm/msm_iommu.c +0
-0@@ -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);
drivers/gpu/drm/msm/msm_iommu.c static int msm_iommu_attach(struct msm_mmu *mmu, const char * const *names, · line 39
Quoted 10 lines
drivers/gpu/drm/msm/msm_iommu.c +10
-0+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?
Quoted 30 lines
drivers/gpu/drm/msm/msm_iommu.c +3
-0return 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);drivers/gpu/drm/msm/msm_mmu.h +3
-1@@ -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
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project
Re: [PATCH 5/5] drm/msm/a6xx: Add support for using system cache(LLC)
Quoted 1 line · Jordan Crouse
On Thu, Dec 19, 2019 at 06:44:46PM +0530, Sharat Masetty wrote:
<snip>
Quoted 17 lines
Quoted 4 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4
-0+ +/* +* CNTL1 is used to specify SCID for (CP, TP, VFD, CCU and UBWC +* FLAG cache) GPU blocks. This value will be passed along with drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 879
Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +10
-0+* 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:
Quoted 6 lines
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
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project
Re: [PATCH 5/5] drm/msm/a6xx: Add support for using system cache(LLC)
In reply to Jordan CrouseQuoted 249 lines · Jordan Crouse
Quoted 41 lines · Sharat Masetty
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_CACHEis 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_CACHEis 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 <[email protected]>--- 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(-)drivers/gpu/drm/msm/adreno/a6xx_gpu.c +1
-0@@ -9,6 +9,7 @@#include "a6xx_gmu.xml.h" #include <linux/devfreq.h> +#include <linux/soc/qcom/llcc-qcom.h> #define GPU_PAS_ID 13drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void · line 787
Quoted 6 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +2
-0@@ -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 5As I mention below, I'm not sure if we need these
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 794
Quoted 8 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +7
-0+#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_SCIDGENMASK(24, 0)
#defineA6XX_GPUHTW_LLC_SCIDGENMASK(29, 25)drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 795
Quoted 1 line
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +1
-0+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.
Quoted 13 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +13
-0+ 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) +{drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 818
Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +10
-0+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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 821
Quoted 3 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +3
-0+ +/* 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);drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 826
Quoted 5 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +5
-0+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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 830
Quoted 4 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4
-0+ +/* 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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 836
Quoted 6 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +6
-0+} + +static void a6xx_llc_slices_destroy (struct a6xx_llc *llc) +{ +if (llc->mmio) +iounmap (llc->mmio);
msm_ioremapreturns adevm_managed resource, so do not useiounmap()to free it. Bets to just leave it and let the gpu device handle it when it goes boom.drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 842
Quoted 6 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +6
-0+ +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.drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 846
Quoted 4 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4
-0+ struct a6xx_llc *llc) +{ + llc->mmio =msm_ioremap (pdev,"cx_mem" ,"gpu_cx" ); +if (IS_ERR_OR_NULL (llc->mmio))
msm_ioremapcan not return NULL.Quoted 9 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +9
-0+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 drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 865
Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +10
-0+* 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.
Quoted 4 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4
-0+ +/* +* CNTL1 is used to specify SCID for (CP, TP, VFD, CCU and UBWC +* FLAG cache) GPU blocks. This value will be passed along with drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 879
Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +10
-0+* 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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 891
Quoted 12 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +12
-0+ } + +/* +* 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, )Quoted 38 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +9
-0+ +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;
Quoted 5 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c struct msm_gpu *a6xx_gpu_init(struct drm_device *dev) · line 1045
Quoted 3 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +2
-1- 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_initpasses 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?
Quoted 90 lines
Quoted 11 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +2
-0+ 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));drivers/gpu/drm/msm/adreno/a6xx_gpu.c +0
-0return ERR_PTR (ret);drivers/gpu/drm/msm/adreno/a6xx_gpu.h +0
-0drivers/gpu/drm/msm/adreno/a6xx_gpu.h line 19
Quoted 8 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.h +5
-0@@ -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.
Quoted 23 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.h +4
-0+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)drivers/gpu/drm/msm/msm_iommu.c +0
-0@@ -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);drivers/gpu/drm/msm/msm_iommu.c static int msm_iommu_attach(struct msm_mmu *mmu, · line 40
Quoted 10 lines
drivers/gpu/drm/msm/msm_iommu.c +10
-0+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?
Quoted 31 lines
drivers/gpu/drm/msm/msm_iommu.c +3
-0return 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);drivers/gpu/drm/msm/msm_mmu.h +3
-1@@ -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
Re: [Freedreno] [PATCH 5/5] drm/msm/a6xx: Add support for using system cache(LLC)
Quoted 271 lines · On Fri, Dec 20, 2019 at 03:40:59PM +0530, [email protected] wrote:
Quoted 264 lines · Jordan Crouse
Quoted 47 lines · Sharat Masetty
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_CACHEis 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_CACHEis 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 <[email protected]>---
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(-)drivers/gpu/drm/msm/adreno/a6xx_gpu.c +1
-0@@ -9,6 +9,7 @@#include "a6xx_gmu.xml.h" #include <linux/devfreq.h> +#include <linux/soc/qcom/llcc-qcom.h> #define GPU_PAS_ID 13drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void · line 787
Quoted 6 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +2
-0@@ -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 5As I mention below, I'm not sure if we need these
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 794
Quoted 8 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +7
-0+#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_SCIDGENMASK(24, 0)
#defineA6XX_GPUHTW_LLC_SCIDGENMASK(29, 25)drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 795
Quoted 1 line
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +1
-0+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.
Quoted 13 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +13
-0+ 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) +{drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 818
Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +10
-0+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.drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 821
Quoted 3 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +3
-0+ +/* 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);drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 826
Quoted 5 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +5
-0+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.drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 830
Quoted 4 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4
-0+ +/* 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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 836
Quoted 6 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +6
-0+} + +static void a6xx_llc_slices_destroy (struct a6xx_llc *llc) +{ +if (llc->mmio) +iounmap (llc->mmio);
msm_ioremapreturns adevm_managed resource, so do not useiounmap()to free
it. Bets to just leave it and let the gpu device handle it when it goes boom.drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 842
Quoted 6 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +6
-0+ +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.drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 846
Quoted 4 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4
-0+ struct a6xx_llc *llc) +{ + llc->mmio =msm_ioremap (pdev,"cx_mem" ,"gpu_cx" ); +if (IS_ERR_OR_NULL (llc->mmio))
msm_ioremapcan not return NULL.Quoted 9 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +9
-0+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 drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 865
Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +10
-0+* 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.Quoted 4 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4
-0+ +/* +* CNTL1 is used to specify SCID for (CP, TP, VFD, CCU and UBWC +* FLAG cache) GPU blocks. This value will be passed along with drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 879
Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +10
-0+* 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.
drivers/gpu/drm/msm/adreno/a6xx_gpu.c static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) · line 891
Quoted 12 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +12
-0+ } + +/* +* 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, useFIELD_SET(A6XX_GPUHTW_LLC_SCID, )Quoted 38 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4
-0+ +return 0 ; +} +static int a6xx_pm_resume (struct msm_gpu *gpu){ struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);drivers/gpu/drm/msm/adreno/a6xx_gpu.c +5
-0@@ -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;Quoted 5 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c struct msm_gpu *a6xx_gpu_init(struct drm_device *dev) · line 1045
Quoted 3 lines
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +2
-1- 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_initpasses or fails.
Quoted 7 lines
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
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project