From d434b8d96d2f632607a78f46828186a98c37183d Mon Sep 17 00:00:00 2001 From: Shaoyun Liu Date: Mon, 24 Apr 2017 10:57:54 -0400 Subject: [PATCH 1666/4131] drm/amdgpu: Implement new invalidate tlb in amdgpu 1. Add invalidate tlb interface in kfd2kgd. 2. Switch to use command package for tlb invalidation on asic newer than v8 when kiq is enabled. Change-Id: Ia6c04ca7a3a57423cc1ab95c42d0787705957151 Signed-off-by: Shaoyun Liu Conflicts: drivers/gpu/drm/amd/amdgpu/soc15d.h drivers/gpu/drm/amd/amdgpu/vid.h --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c | 23 ++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c | 52 ++++++++++++++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c | 51 ++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/vid.h | 4 ++ drivers/gpu/drm/amd/include/kgd_kfd_interface.h | 3 ++ 5 files changed, 132 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c index 073a500..2abf945 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c @@ -144,6 +144,7 @@ static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd, uint8_t vmid); static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd, uint8_t vmid); static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid); +static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid); static void set_num_of_requests(struct kgd_dev *dev, uint8_t num_of_req); static int alloc_memory_of_scratch(struct kgd_dev *kgd, uint64_t va, uint32_t vmid); @@ -211,6 +212,7 @@ static const struct kfd2kgd_calls kfd2kgd = { get_atc_vmid_pasid_mapping_valid, .read_vmid_from_vmfault_reg = read_vmid_from_vmfault_reg, .write_vmid_invalidate_request = write_vmid_invalidate_request, + .invalidate_tlbs = invalidate_tlbs, .alloc_memory_of_gpu = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu, .free_memory_of_gpu = amdgpu_amdkfd_gpuvm_free_memory_of_gpu, .map_memory_to_gpu = amdgpu_amdkfd_gpuvm_map_memory_to_gpu, @@ -810,6 +812,27 @@ static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid) WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); } +static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid) +{ + struct amdgpu_device *adev = (struct amdgpu_device *) kgd; + int vmid; + + for (vmid = 0; vmid < 16; vmid++) { + if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) + continue; + if (RREG32(mmATC_VMID0_PASID_MAPPING + vmid) & + ATC_VMID0_PASID_MAPPING__VALID_MASK) { + if ((RREG32(mmATC_VMID0_PASID_MAPPING + vmid) & + ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) { + WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); + break; + } + } + } + + return 0; +} + static int write_config_static_mem(struct kgd_dev *kgd, bool swizzle_enable, uint8_t element_size, uint8_t index_stride, uint8_t mtype) { diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c index d92dce7..452bebd 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c @@ -104,7 +104,6 @@ static int kgd_hqd_destroy(struct kgd_dev *kgd, uint32_t queue_id); static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd, unsigned int utimeout); -static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid); static int kgd_address_watch_disable(struct kgd_dev *kgd); static int kgd_address_watch_execute(struct kgd_dev *kgd, unsigned int watch_point_id, @@ -131,6 +130,7 @@ static int write_config_static_mem(struct kgd_dev *kgd, bool swizzle_enable, uint8_t element_size, uint8_t index_stride, uint8_t mtype); static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid, uint32_t page_table_base); +static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid); /* Because of REG_GET_FIELD() being used, we put this function in the * asic specific file. @@ -190,6 +190,7 @@ static const struct kfd2kgd_calls kfd2kgd = { .get_atc_vmid_pasid_mapping_valid = get_atc_vmid_pasid_mapping_valid, .write_vmid_invalidate_request = write_vmid_invalidate_request, + .invalidate_tlbs = invalidate_tlbs, .alloc_memory_of_gpu = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu, .free_memory_of_gpu = amdgpu_amdkfd_gpuvm_free_memory_of_gpu, .map_memory_to_gpu = amdgpu_amdkfd_gpuvm_map_memory_to_gpu, @@ -757,6 +758,55 @@ static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid) WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); } +static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid) +{ + signed long r; + struct fence *f; + struct amdgpu_ring *ring = &adev->gfx.kiq.ring; + + mutex_lock(&adev->virt.lock_kiq); + amdgpu_ring_alloc(ring, 12); /* fence + invalidate_tlbs package*/ + amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0)); + amdgpu_ring_write(ring, + PACKET3_INVALIDATE_TLBS_DST_SEL(1) | + PACKET3_INVALIDATE_TLBS_PASID(pasid)); + amdgpu_fence_emit(ring, &f); + amdgpu_ring_commit(ring); + mutex_unlock(&adev->virt.lock_kiq); + + r = fence_wait(f, false); + if (r) + DRM_ERROR("wait for kiq fence error: %ld.\n", r); + fence_put(f); + + return r; +} + +static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid) +{ + struct amdgpu_device *adev = (struct amdgpu_device *) kgd; + int vmid; + struct amdgpu_ring *ring = &adev->gfx.kiq.ring; + + if (ring->ready) + return invalidate_tlbs_with_kiq(adev, pasid); + + for (vmid = 0; vmid < 16; vmid++) { + if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) + continue; + if (RREG32(mmATC_VMID0_PASID_MAPPING + vmid) & + ATC_VMID0_PASID_MAPPING__VALID_MASK) { + if ((RREG32(mmATC_VMID0_PASID_MAPPING + vmid) & + ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) { + WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); + break; + } + } + } + + return 0; +} + static int kgd_address_watch_disable(struct kgd_dev *kgd) { struct amdgpu_device *adev = get_amdgpu_device(kgd); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c index 796dc44..a2fe8be 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c @@ -46,6 +46,7 @@ #include "soc15_common.h" #include "v9_structs.h" #include "soc15.h" +#include "soc15d.h" /* HACK: MMHUB and GC both have VM-related register with the same * names but different offsets. Define the MMHUB register we need here @@ -168,6 +169,7 @@ static int write_config_static_mem(struct kgd_dev *kgd, bool swizzle_enable, uint8_t element_size, uint8_t index_stride, uint8_t mtype); static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid, uint32_t page_table_base); +static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid); /* Because of REG_GET_FIELD() being used, we put this function in the * asic specific file. @@ -234,6 +236,7 @@ static const struct kfd2kgd_calls kfd2kgd = { .get_atc_vmid_pasid_mapping_valid = get_atc_vmid_pasid_mapping_valid, .write_vmid_invalidate_request = write_vmid_invalidate_request, + .invalidate_tlbs = invalidate_tlbs, .alloc_memory_of_gpu = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu, .free_memory_of_gpu = amdgpu_amdkfd_gpuvm_free_memory_of_gpu, .map_memory_to_gpu = amdgpu_amdkfd_gpuvm_map_memory_to_gpu, @@ -933,6 +936,54 @@ static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid) cpu_relax(); } +static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid) +{ + signed long r; + struct fence *f; + struct amdgpu_ring *ring = &adev->gfx.kiq.ring; + + mutex_lock(&adev->virt.lock_kiq); + amdgpu_ring_alloc(ring, 12); /* fence + invalidate_tlbs package*/ + amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0)); + amdgpu_ring_write(ring, + PACKET3_INVALIDATE_TLBS_DST_SEL(1) | + PACKET3_INVALIDATE_TLBS_PASID(pasid)); + amdgpu_fence_emit(ring, &f); + amdgpu_ring_commit(ring); + mutex_unlock(&adev->virt.lock_kiq); + + r = fence_wait(f, false); + if (r) + DRM_ERROR("wait for kiq fence error: %ld.\n", r); + fence_put(f); + + return r; +} + +static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid) +{ + struct amdgpu_device *adev = (struct amdgpu_device *) kgd; + int vmid; + struct amdgpu_ring *ring = &adev->gfx.kiq.ring; + + if (ring->ready) + return invalidate_tlbs_with_kiq(adev, pasid); + + for (vmid = 0; vmid < 16; vmid++) { + if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) + continue; + if (get_atc_vmid_pasid_mapping_valid(kgd, vmid)) { + if (get_atc_vmid_pasid_mapping_pasid(kgd, vmid) + == pasid) { + write_vmid_invalidate_request(kgd, vmid); + break; + } + } + } + + return 0; +} + static int kgd_address_watch_disable(struct kgd_dev *kgd) { struct amdgpu_device *adev = get_amdgpu_device(kgd); diff --git a/drivers/gpu/drm/amd/amdgpu/vid.h b/drivers/gpu/drm/amd/amdgpu/vid.h index 9806900..d09592a 100755 --- a/drivers/gpu/drm/amd/amdgpu/vid.h +++ b/drivers/gpu/drm/amd/amdgpu/vid.h @@ -369,6 +369,10 @@ * x=0: tmz_begin * x=1: tmz_end */ +#define PACKET3_INVALIDATE_TLBS 0x98 +# define PACKET3_INVALIDATE_TLBS_DST_SEL(x) ((x) << 0) +# define PACKET3_INVALIDATE_TLBS_PASID(x) ((x) << 5) + #define PACKET3_SET_RESOURCES 0xA0 /* 1. header * 2. CONTROL diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h index cff1ee6..132a801 100644 --- a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h +++ b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h @@ -335,6 +335,9 @@ struct kfd2kgd_calls { uint32_t (*read_vmid_from_vmfault_reg)(struct kgd_dev *kgd); void (*write_vmid_invalidate_request)(struct kgd_dev *kgd, uint8_t vmid); + + int (*invalidate_tlbs)(struct kgd_dev *kgd, uint16_t pasid); + int (*alloc_memory_of_gpu)(struct kgd_dev *kgd, uint64_t va, uint64_t size, void *vm, struct kgd_mem **mem, uint64_t *offset, -- 2.7.4