1 files changed, 164 insertions, 81 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index 1580ec60b89f..fe1d7368c1e6 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -26,22 +26,28 @@
 #include "amdgpu.h"
 #include "amdgpu_gfx.h"
 #include <linux/module.h>
-
-const struct kgd2kfd_calls *kgd2kfd;
+#include <linux/dma-buf.h>
 
 static const unsigned int compute_vmid_bitmap = 0xFF00;
 
+/* Total memory size in system memory and all GPU VRAM. Used to
+ * estimate worst case amount of memory to reserve for page tables
+ */
+uint64_t amdgpu_amdkfd_total_mem_size;
+
 int amdgpu_amdkfd_init(void)
 {
+	struct sysinfo si;
 	int ret;
 
+	si_meminfo(&si);
+	amdgpu_amdkfd_total_mem_size = si.totalram - si.totalhigh;
+	amdgpu_amdkfd_total_mem_size *= si.mem_unit;
+
 #ifdef CONFIG_HSA_AMD
-	ret = kgd2kfd_init(KFD_INTERFACE_VERSION, &kgd2kfd);
-	if (ret)
-		kgd2kfd = NULL;
+	ret = kgd2kfd_init();
 	amdgpu_amdkfd_gpuvm_init_mem_limits();
 #else
-	kgd2kfd = NULL;
 	ret = -ENOENT;
 #endif
 
@@ -50,17 +56,13 @@ int amdgpu_amdkfd_init(void)
 
 void amdgpu_amdkfd_fini(void)
 {
-	if (kgd2kfd)
-		kgd2kfd->exit();
+	kgd2kfd_exit();
 }
 
 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
 {
 	const struct kfd2kgd_calls *kfd2kgd;
 
-	if (!kgd2kfd)
-		return;
-
 	switch (adev->asic_type) {
 #ifdef CONFIG_DRM_AMDGPU_CIK
 	case CHIP_KAVERI:
@@ -73,9 +75,11 @@ void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
 	case CHIP_FIJI:
 	case CHIP_POLARIS10:
 	case CHIP_POLARIS11:
+	case CHIP_POLARIS12:
 		kfd2kgd = amdgpu_amdkfd_gfx_8_0_get_functions();
 		break;
 	case CHIP_VEGA10:
+	case CHIP_VEGA12:
 	case CHIP_VEGA20:
 	case CHIP_RAVEN:
 		kfd2kgd = amdgpu_amdkfd_gfx_9_0_get_functions();
@@ -85,8 +89,11 @@ void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
 		return;
 	}
 
-	adev->kfd = kgd2kfd->probe((struct kgd_dev *)adev,
-				   adev->pdev, kfd2kgd);
+	adev->kfd.dev = kgd2kfd_probe((struct kgd_dev *)adev,
+				      adev->pdev, kfd2kgd);
+
+	if (adev->kfd.dev)
+		amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size;
 }
 
 /**
@@ -124,9 +131,10 @@ static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
 
 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
 {
-	int i, n;
+	int i;
 	int last_valid_bit;
-	if (adev->kfd) {
+
+	if (adev->kfd.dev) {
 		struct kgd2kfd_shared_resources gpu_resources = {
 			.compute_vmid_bitmap = compute_vmid_bitmap,
 			.num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec,
@@ -134,7 +142,9 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
 			.gpuvm_size = min(adev->vm_manager.max_pfn
 					  << AMDGPU_GPU_PAGE_SHIFT,
 					  AMDGPU_GMC_HOLE_START),
-			.drm_render_minor = adev->ddev->render->index
+			.drm_render_minor = adev->ddev->render->index,
+			.sdma_doorbell_idx = adev->doorbell_index.sdma_engine,
+
 		};
 
 		/* this is going to have a few of the MSBs set that we need to
@@ -144,7 +154,7 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
 				  KGD_MAX_QUEUES);
 
 		/* remove the KIQ bit as well */
-		if (adev->gfx.kiq.ring.ready)
+		if (adev->gfx.kiq.ring.sched.ready)
 			clear_bit(amdgpu_gfx_queue_to_bit(adev,
 							  adev->gfx.kiq.ring.me - 1,
 							  adev->gfx.kiq.ring.pipe,
@@ -164,78 +174,52 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
 				&gpu_resources.doorbell_aperture_size,
 				&gpu_resources.doorbell_start_offset);
 
-		if (adev->asic_type < CHIP_VEGA10) {
-			kgd2kfd->device_init(adev->kfd, &gpu_resources);
-			return;
-		}
-
-		n = (adev->asic_type < CHIP_VEGA20) ? 2 : 8;
-
-		for (i = 0; i < n; i += 2) {
-			/* On SOC15 the BIF is involved in routing
-			 * doorbells using the low 12 bits of the
-			 * address. Communicate the assignments to
-			 * KFD. KFD uses two doorbell pages per
-			 * process in case of 64-bit doorbells so we
-			 * can use each doorbell assignment twice.
-			 */
-			if (adev->asic_type == CHIP_VEGA10) {
-				gpu_resources.sdma_doorbell[0][i] =
-					AMDGPU_VEGA10_DOORBELL64_sDMA_ENGINE0 + (i >> 1);
-				gpu_resources.sdma_doorbell[0][i+1] =
-					AMDGPU_VEGA10_DOORBELL64_sDMA_ENGINE0 + 0x200 + (i >> 1);
-				gpu_resources.sdma_doorbell[1][i] =
-					AMDGPU_VEGA10_DOORBELL64_sDMA_ENGINE1 + (i >> 1);
-				gpu_resources.sdma_doorbell[1][i+1] =
-					AMDGPU_VEGA10_DOORBELL64_sDMA_ENGINE1 + 0x200 + (i >> 1);
-			} else {
-				gpu_resources.sdma_doorbell[0][i] =
-					AMDGPU_DOORBELL64_sDMA_ENGINE0 + (i >> 1);
-				gpu_resources.sdma_doorbell[0][i+1] =
-					AMDGPU_DOORBELL64_sDMA_ENGINE0 + 0x200 + (i >> 1);
-				gpu_resources.sdma_doorbell[1][i] =
-					AMDGPU_DOORBELL64_sDMA_ENGINE1 + (i >> 1);
-				gpu_resources.sdma_doorbell[1][i+1] =
-					AMDGPU_DOORBELL64_sDMA_ENGINE1 + 0x200 + (i >> 1);
-			}
-		}
-		/* Doorbells 0x0e0-0ff and 0x2e0-2ff are reserved for
-		 * SDMA, IH and VCN. So don't use them for the CP.
+		/* Since SOC15, BIF starts to statically use the
+		 * lower 12 bits of doorbell addresses for routing
+		 * based on settings in registers like
+		 * SDMA0_DOORBELL_RANGE etc..
+		 * In order to route a doorbell to CP engine, the lower
+		 * 12 bits of its address has to be outside the range
+		 * set for SDMA, VCN, and IH blocks.
 		 */
-		gpu_resources.reserved_doorbell_mask = 0x1e0;
-		gpu_resources.reserved_doorbell_val  = 0x0e0;
+		if (adev->asic_type >= CHIP_VEGA10) {
+			gpu_resources.non_cp_doorbells_start =
+					adev->doorbell_index.first_non_cp;
+			gpu_resources.non_cp_doorbells_end =
+					adev->doorbell_index.last_non_cp;
+		}
 
-		kgd2kfd->device_init(adev->kfd, &gpu_resources);
+		kgd2kfd_device_init(adev->kfd.dev, &gpu_resources);
 	}
 }
 
 void amdgpu_amdkfd_device_fini(struct amdgpu_device *adev)
 {
-	if (adev->kfd) {
-		kgd2kfd->device_exit(adev->kfd);
-		adev->kfd = NULL;
+	if (adev->kfd.dev) {
+		kgd2kfd_device_exit(adev->kfd.dev);
+		adev->kfd.dev = NULL;
 	}
 }
 
 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
 		const void *ih_ring_entry)
 {
-	if (adev->kfd)
-		kgd2kfd->interrupt(adev->kfd, ih_ring_entry);
+	if (adev->kfd.dev)
+		kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry);
 }
 
 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev)
 {
-	if (adev->kfd)
-		kgd2kfd->suspend(adev->kfd);
+	if (adev->kfd.dev)
+		kgd2kfd_suspend(adev->kfd.dev);
 }
 
 int amdgpu_amdkfd_resume(struct amdgpu_device *adev)
 {
 	int r = 0;
 
-	if (adev->kfd)
-		r = kgd2kfd->resume(adev->kfd);
+	if (adev->kfd.dev)
+		r = kgd2kfd_resume(adev->kfd.dev);
 
 	return r;
 }
@@ -244,8 +228,8 @@ int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev)
 {
 	int r = 0;
 
-	if (adev->kfd)
-		r = kgd2kfd->pre_reset(adev->kfd);
+	if (adev->kfd.dev)
+		r = kgd2kfd_pre_reset(adev->kfd.dev);
 
 	return r;
 }
@@ -254,8 +238,8 @@ int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev)
 {
 	int r = 0;
 
-	if (adev->kfd)
-		r = kgd2kfd->post_reset(adev->kfd);
+	if (adev->kfd.dev)
+		r = kgd2kfd_post_reset(adev->kfd.dev);
 
 	return r;
 }
@@ -268,9 +252,9 @@ void amdgpu_amdkfd_gpu_reset(struct kgd_dev *kgd)
 		amdgpu_device_gpu_recover(adev, NULL);
 }
 
-int alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
-			void **mem_obj, uint64_t *gpu_addr,
-			void **cpu_ptr, bool mqd_gfx9)
+int amdgpu_amdkfd_alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
+				void **mem_obj, uint64_t *gpu_addr,
+				void **cpu_ptr, bool mqd_gfx9)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 	struct amdgpu_bo *bo = NULL;
@@ -340,7 +324,7 @@ allocate_mem_reserve_bo_failed:
 	return r;
 }
 
-void free_gtt_mem(struct kgd_dev *kgd, void *mem_obj)
+void amdgpu_amdkfd_free_gtt_mem(struct kgd_dev *kgd, void *mem_obj)
 {
 	struct amdgpu_bo *bo = (struct amdgpu_bo *) mem_obj;
 
@@ -351,8 +335,8 @@ void free_gtt_mem(struct kgd_dev *kgd, void *mem_obj)
 	amdgpu_bo_unref(&(bo));
 }
 
-void get_local_mem_info(struct kgd_dev *kgd,
-			struct kfd_local_mem_info *mem_info)
+void amdgpu_amdkfd_get_local_mem_info(struct kgd_dev *kgd,
+				      struct kfd_local_mem_info *mem_info)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 	uint64_t address_mask = adev->dev->dma_mask ? ~*adev->dev->dma_mask :
@@ -383,7 +367,7 @@ void get_local_mem_info(struct kgd_dev *kgd,
 		mem_info->mem_clk_max = 100;
 }
 
-uint64_t get_gpu_clock_counter(struct kgd_dev *kgd)
+uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct kgd_dev *kgd)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 
@@ -392,7 +376,7 @@ uint64_t get_gpu_clock_counter(struct kgd_dev *kgd)
 	return 0;
 }
 
-uint32_t get_max_engine_clock_in_mhz(struct kgd_dev *kgd)
+uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct kgd_dev *kgd)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 
@@ -405,7 +389,7 @@ uint32_t get_max_engine_clock_in_mhz(struct kgd_dev *kgd)
 		return 100;
 }
 
-void get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info)
+void amdgpu_amdkfd_get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 	struct amdgpu_cu_info acu_info = adev->gfx.cu_info;
@@ -428,6 +412,62 @@ void get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info)
 	cu_info->lds_size = acu_info.lds_size;
 }
 
+int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd,
+				  struct kgd_dev **dma_buf_kgd,
+				  uint64_t *bo_size, void *metadata_buffer,
+				  size_t buffer_size, uint32_t *metadata_size,
+				  uint32_t *flags)
+{
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
+	struct dma_buf *dma_buf;
+	struct drm_gem_object *obj;
+	struct amdgpu_bo *bo;
+	uint64_t metadata_flags;
+	int r = -EINVAL;
+
+	dma_buf = dma_buf_get(dma_buf_fd);
+	if (IS_ERR(dma_buf))
+		return PTR_ERR(dma_buf);
+
+	if (dma_buf->ops != &amdgpu_dmabuf_ops)
+		/* Can't handle non-graphics buffers */
+		goto out_put;
+
+	obj = dma_buf->priv;
+	if (obj->dev->driver != adev->ddev->driver)
+		/* Can't handle buffers from different drivers */
+		goto out_put;
+
+	adev = obj->dev->dev_private;
+	bo = gem_to_amdgpu_bo(obj);
+	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
+				    AMDGPU_GEM_DOMAIN_GTT)))
+		/* Only VRAM and GTT BOs are supported */
+		goto out_put;
+
+	r = 0;
+	if (dma_buf_kgd)
+		*dma_buf_kgd = (struct kgd_dev *)adev;
+	if (bo_size)
+		*bo_size = amdgpu_bo_size(bo);
+	if (metadata_size)
+		*metadata_size = bo->metadata_size;
+	if (metadata_buffer)
+		r = amdgpu_bo_get_metadata(bo, metadata_buffer, buffer_size,
+					   metadata_size, &metadata_flags);
+	if (flags) {
+		*flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
+			ALLOC_MEM_FLAGS_VRAM : ALLOC_MEM_FLAGS_GTT;
+
+		if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
+			*flags |= ALLOC_MEM_FLAGS_PUBLIC;
+	}
+
+out_put:
+	dma_buf_put(dma_buf);
+	return r;
+}
+
 uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
@@ -510,7 +550,7 @@ void amdgpu_amdkfd_set_compute_idle(struct kgd_dev *kgd, bool idle)
 
 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid)
 {
-	if (adev->kfd) {
+	if (adev->kfd.dev) {
 		if ((1 << vmid) & compute_vmid_bitmap)
 			return true;
 	}
@@ -524,7 +564,7 @@ bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
 	return false;
 }
 
-void amdgpu_amdkfd_unreserve_system_memory_limit(struct amdgpu_bo *bo)
+void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo)
 {
 }
 
@@ -557,4 +597,47 @@ struct kfd2kgd_calls *amdgpu_amdkfd_gfx_9_0_get_functions(void)
 {
 	return NULL;
 }
+
+struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev,
+			      const struct kfd2kgd_calls *f2g)
+{
+	return NULL;
+}
+
+bool kgd2kfd_device_init(struct kfd_dev *kfd,
+			 const struct kgd2kfd_shared_resources *gpu_resources)
+{
+	return false;
+}
+
+void kgd2kfd_device_exit(struct kfd_dev *kfd)
+{
+}
+
+void kgd2kfd_exit(void)
+{
+}
+
+void kgd2kfd_suspend(struct kfd_dev *kfd)
+{
+}
+
+int kgd2kfd_resume(struct kfd_dev *kfd)
+{
+	return 0;
+}
+
+int kgd2kfd_pre_reset(struct kfd_dev *kfd)
+{
+	return 0;
+}
+
+int kgd2kfd_post_reset(struct kfd_dev *kfd)
+{
+	return 0;
+}
+
+void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
+{
+}
 #endif