drm/amdgpu: Add GPUVM memory management functions for KFD

v2:
* Removed unused flags from struct kgd_mem
* Updated some comments
* Added a check to unmap_memory_from_gpu whether BO was mapped

v3: add mutex_destroy in relevant places

Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Oded Gabbay <oded.gabbay@gmail.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>

MAINTAINERS

@@ -767,6 +767,7 @@ F:	drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
 F:	drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c
 F:	drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c
 F:	drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
+F:	drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
 F:	drivers/gpu/drm/amd/amdkfd/
 F:	drivers/gpu/drm/amd/include/cik_structs.h
 F:	drivers/gpu/drm/amd/include/kgd_kfd_interface.h

drivers/gpu/drm/amd/amdgpu/Makefile

@@ -130,6 +130,7 @@ amdgpu-y += \
 amdgpu-y += \
 	 amdgpu_amdkfd.o \
 	 amdgpu_amdkfd_fence.o \
+	 amdgpu_amdkfd_gpuvm.o \
 	 amdgpu_amdkfd_gfx_v8.o
 
 # add cgs

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c

@@ -58,6 +58,7 @@ int amdgpu_amdkfd_init(void)
 #else
 	ret = -ENOENT;
 #endif
+	amdgpu_amdkfd_gpuvm_init_mem_limits();
 
 	return ret;
 }

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h

@@ -28,15 +28,41 @@
 #include <linux/types.h>
 #include <linux/mmu_context.h>
 #include <kgd_kfd_interface.h>
+#include <drm/ttm/ttm_execbuf_util.h>
+#include "amdgpu_sync.h"
+#include "amdgpu_vm.h"
 
 extern const struct kgd2kfd_calls *kgd2kfd;
 
 struct amdgpu_device;
 
+struct kfd_bo_va_list {
+	struct list_head bo_list;
+	struct amdgpu_bo_va *bo_va;
+	void *kgd_dev;
+	bool is_mapped;
+	uint64_t va;
+	uint64_t pte_flags;
+};
+
 struct kgd_mem {
+	struct mutex lock;
 	struct amdgpu_bo *bo;
-	uint64_t gpu_addr;
+	struct list_head bo_va_list;
-	void *cpu_ptr;
+	/* protected by amdkfd_process_info.lock */
+	struct ttm_validate_buffer validate_list;
+	struct ttm_validate_buffer resv_list;
+	uint32_t domain;
+	unsigned int mapped_to_gpu_memory;
+	uint64_t va;
+
+	uint32_t mapping_flags;
+
+	struct amdkfd_process_info *process_info;
+
+	struct amdgpu_sync sync;
+
+	bool aql_queue;
 };
 
 /* KFD Memory Eviction */
@@ -52,6 +78,41 @@ struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm);
 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f);
 
+struct amdkfd_process_info {
+	/* List head of all VMs that belong to a KFD process */
+	struct list_head vm_list_head;
+	/* List head for all KFD BOs that belong to a KFD process. */
+	struct list_head kfd_bo_list;
+	/* Lock to protect kfd_bo_list */
+	struct mutex lock;
+
+	/* Number of VMs */
+	unsigned int n_vms;
+	/* Eviction Fence */
+	struct amdgpu_amdkfd_fence *eviction_fence;
+};
+
+/* struct amdkfd_vm -
+ * For Memory Eviction KGD requires a mechanism to keep track of all KFD BOs
+ * belonging to a KFD process. All the VMs belonging to the same process point
+ * to the same amdkfd_process_info.
+ */
+struct amdkfd_vm {
+	/* Keep base as the first parameter for pointer compatibility between
+	 * amdkfd_vm and amdgpu_vm.
+	 */
+	struct amdgpu_vm base;
+
+	/* List node in amdkfd_process_info.vm_list_head*/
+	struct list_head vm_list_node;
+
+	struct amdgpu_device *adev;
+	/* Points to the KFD process VM info*/
+	struct amdkfd_process_info *process_info;
+
+	uint64_t pd_phys_addr;
+};
+
 int amdgpu_amdkfd_init(void);
 void amdgpu_amdkfd_fini(void);
 
@@ -96,4 +157,30 @@ uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd);
 		valid;							\
 	})
 
+/* GPUVM API */
+int amdgpu_amdkfd_gpuvm_create_process_vm(struct kgd_dev *kgd, void **vm,
+					  void **process_info,
+					  struct dma_fence **ef);
+void amdgpu_amdkfd_gpuvm_destroy_process_vm(struct kgd_dev *kgd, void *vm);
+uint32_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *vm);
+int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
+		struct kgd_dev *kgd, uint64_t va, uint64_t size,
+		void *vm, struct kgd_mem **mem,
+		uint64_t *offset, uint32_t flags);
+int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
+		struct kgd_dev *kgd, struct kgd_mem *mem);
+int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
+		struct kgd_dev *kgd, struct kgd_mem *mem, void *vm);
+int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
+		struct kgd_dev *kgd, struct kgd_mem *mem, void *vm);
+int amdgpu_amdkfd_gpuvm_sync_memory(
+		struct kgd_dev *kgd, struct kgd_mem *mem, bool intr);
+int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd,
+		struct kgd_mem *mem, void **kptr, uint64_t *size);
+int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info,
+					    struct dma_fence **ef);
+
+void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
+void amdgpu_amdkfd_unreserve_system_memory_limit(struct amdgpu_bo *bo);
+
 #endif /* AMDGPU_AMDKFD_H_INCLUDED */

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c

@@ -143,6 +143,10 @@ static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type);
 static void set_scratch_backing_va(struct kgd_dev *kgd,
 					uint64_t va, uint32_t vmid);
+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);
+static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
 
 /* Because of REG_GET_FIELD() being used, we put this function in the
  * asic specific file.
@@ -199,7 +203,20 @@ static const struct kfd2kgd_calls kfd2kgd = {
 	.set_scratch_backing_va = set_scratch_backing_va,
 	.get_tile_config = get_tile_config,
 	.get_cu_info = get_cu_info,
-	.get_vram_usage = amdgpu_amdkfd_get_vram_usage
+	.get_vram_usage = amdgpu_amdkfd_get_vram_usage,
+	.create_process_vm = amdgpu_amdkfd_gpuvm_create_process_vm,
+	.destroy_process_vm = amdgpu_amdkfd_gpuvm_destroy_process_vm,
+	.get_process_page_dir = amdgpu_amdkfd_gpuvm_get_process_page_dir,
+	.set_vm_context_page_table_base = set_vm_context_page_table_base,
+	.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,
+	.unmap_memory_to_gpu = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu,
+	.sync_memory = amdgpu_amdkfd_gpuvm_sync_memory,
+	.map_gtt_bo_to_kernel = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel,
+	.restore_process_bos = amdgpu_amdkfd_gpuvm_restore_process_bos,
+	.invalidate_tlbs = invalidate_tlbs,
+	.invalidate_tlbs_vmid = invalidate_tlbs_vmid,
 };
 
 struct kfd2kgd_calls *amdgpu_amdkfd_gfx_7_get_functions(void)
@@ -855,3 +872,50 @@ static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type)
 	return hdr->common.ucode_version;
 }
 
+static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
+			uint32_t page_table_base)
+{
+	struct amdgpu_device *adev = get_amdgpu_device(kgd);
+
+	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
+		pr_err("trying to set page table base for wrong VMID\n");
+		return;
+	}
+	WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8, page_table_base);
+}
+
+static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
+{
+	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
+	int vmid;
+	unsigned int tmp;
+
+	for (vmid = 0; vmid < 16; vmid++) {
+		if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
+			continue;
+
+		tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
+		if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) &&
+			(tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) {
+			WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
+			RREG32(mmVM_INVALIDATE_RESPONSE);
+			break;
+		}
+	}
+
+	return 0;
+}
+
+static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
+{
+	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
+
+	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
+		pr_err("non kfd vmid\n");
+		return 0;
+	}
+
+	WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
+	RREG32(mmVM_INVALIDATE_RESPONSE);
+	return 0;
+}

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c

@@ -101,6 +101,10 @@ static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type);
 static void set_scratch_backing_va(struct kgd_dev *kgd,
 					uint64_t va, uint32_t vmid);
+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);
+static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
 
 /* Because of REG_GET_FIELD() being used, we put this function in the
  * asic specific file.
@@ -159,7 +163,20 @@ static const struct kfd2kgd_calls kfd2kgd = {
 	.set_scratch_backing_va = set_scratch_backing_va,
 	.get_tile_config = get_tile_config,
 	.get_cu_info = get_cu_info,
-	.get_vram_usage = amdgpu_amdkfd_get_vram_usage
+	.get_vram_usage = amdgpu_amdkfd_get_vram_usage,
+	.create_process_vm = amdgpu_amdkfd_gpuvm_create_process_vm,
+	.destroy_process_vm = amdgpu_amdkfd_gpuvm_destroy_process_vm,
+	.get_process_page_dir = amdgpu_amdkfd_gpuvm_get_process_page_dir,
+	.set_vm_context_page_table_base = set_vm_context_page_table_base,
+	.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,
+	.unmap_memory_to_gpu = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu,
+	.sync_memory = amdgpu_amdkfd_gpuvm_sync_memory,
+	.map_gtt_bo_to_kernel = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel,
+	.restore_process_bos = amdgpu_amdkfd_gpuvm_restore_process_bos,
+	.invalidate_tlbs = invalidate_tlbs,
+	.invalidate_tlbs_vmid = invalidate_tlbs_vmid,
 };
 
 struct kfd2kgd_calls *amdgpu_amdkfd_gfx_8_0_get_functions(void)
@@ -816,3 +833,51 @@ static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type)
 	/* Only 12 bit in use*/
 	return hdr->common.ucode_version;
 }
+
+static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
+		uint32_t page_table_base)
+{
+	struct amdgpu_device *adev = get_amdgpu_device(kgd);
+
+	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
+		pr_err("trying to set page table base for wrong VMID\n");
+		return;
+	}
+	WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8, page_table_base);
+}
+
+static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
+{
+	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
+	int vmid;
+	unsigned int tmp;
+
+	for (vmid = 0; vmid < 16; vmid++) {
+		if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
+			continue;
+
+		tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
+		if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) &&
+			(tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) {
+			WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
+			RREG32(mmVM_INVALIDATE_RESPONSE);
+			break;
+		}
+	}
+
+	return 0;
+}
+
+static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
+{
+	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
+
+	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
+		pr_err("non kfd vmid %d\n", vmid);
+		return -EINVAL;
+	}
+
+	WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
+	RREG32(mmVM_INVALIDATE_RESPONSE);
+	return 0;
+}

drivers/gpu/drm/amd/amdgpu/amdgpu_object.c

@@ -36,6 +36,7 @@
 #include <drm/drm_cache.h>
 #include "amdgpu.h"
 #include "amdgpu_trace.h"
+#include "amdgpu_amdkfd.h"
 
 static bool amdgpu_need_backup(struct amdgpu_device *adev)
 {
@@ -54,6 +55,9 @@ static void amdgpu_ttm_bo_destroy(struct ttm_buffer_object *tbo)
 	struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
 	struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
 
+	if (bo->kfd_bo)
+		amdgpu_amdkfd_unreserve_system_memory_limit(bo);
+
 	amdgpu_bo_kunmap(bo);
 
 	drm_gem_object_release(&bo->gem_base);

drivers/gpu/drm/amd/amdgpu/amdgpu_object.h

@@ -92,6 +92,8 @@ struct amdgpu_bo {
 		struct list_head	mn_list;
 		struct list_head	shadow_list;
 	};
+
+	struct kgd_mem                  *kfd_bo;
 };
 
 static inline struct amdgpu_bo *ttm_to_amdgpu_bo(struct ttm_buffer_object *tbo)

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

@@ -259,6 +259,13 @@ static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 {
 	struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
 
+	/*
+	 * Don't verify access for KFD BOs. They don't have a GEM
+	 * object associated with them.
+	 */
+	if (abo->kfd_bo)
+		return 0;
+
 	if (amdgpu_ttm_tt_get_usermm(bo->ttm))
 		return -EPERM;
 	return drm_vma_node_verify_access(&abo->gem_base.vma_node,

drivers/gpu/drm/amd/include/kgd_kfd_interface.h

@@ -127,6 +127,25 @@ struct tile_config {
 	uint32_t num_ranks;
 };
 
+
+/*
+ * Allocation flag domains
+ */
+#define ALLOC_MEM_FLAGS_VRAM		(1 << 0)
+#define ALLOC_MEM_FLAGS_GTT		(1 << 1)
+#define ALLOC_MEM_FLAGS_USERPTR		(1 << 2) /* TODO */
+#define ALLOC_MEM_FLAGS_DOORBELL	(1 << 3) /* TODO */
+
+/*
+ * Allocation flags attributes/access options.
+ */
+#define ALLOC_MEM_FLAGS_WRITABLE	(1 << 31)
+#define ALLOC_MEM_FLAGS_EXECUTABLE	(1 << 30)
+#define ALLOC_MEM_FLAGS_PUBLIC		(1 << 29)
+#define ALLOC_MEM_FLAGS_NO_SUBSTITUTE	(1 << 28) /* TODO */
+#define ALLOC_MEM_FLAGS_AQL_QUEUE_MEM	(1 << 27)
+#define ALLOC_MEM_FLAGS_COHERENT	(1 << 26) /* For GFXv9 or later */
+
 /**
  * struct kfd2kgd_calls
  *
@@ -186,6 +205,41 @@ struct tile_config {
  *
  * @get_vram_usage: Returns current VRAM usage
  *
+ * @create_process_vm: Create a VM address space for a given process and GPU
+ *
+ * @destroy_process_vm: Destroy a VM
+ *
+ * @get_process_page_dir: Get physical address of a VM page directory
+ *
+ * @set_vm_context_page_table_base: Program page table base for a VMID
+ *
+ * @alloc_memory_of_gpu: Allocate GPUVM memory
+ *
+ * @free_memory_of_gpu: Free GPUVM memory
+ *
+ * @map_memory_to_gpu: Map GPUVM memory into a specific VM address
+ * space. Allocates and updates page tables and page directories as
+ * needed. This function may return before all page table updates have
+ * completed. This allows multiple map operations (on multiple GPUs)
+ * to happen concurrently. Use sync_memory to synchronize with all
+ * pending updates.
+ *
+ * @unmap_memor_to_gpu: Unmap GPUVM memory from a specific VM address space
+ *
+ * @sync_memory: Wait for pending page table updates to complete
+ *
+ * @map_gtt_bo_to_kernel: Map a GTT BO for kernel access
+ * Pins the BO, maps it to kernel address space. Such BOs are never evicted.
+ * The kernel virtual address remains valid until the BO is freed.
+ *
+ * @restore_process_bos: Restore all BOs that belong to the
+ * process. This is intended for restoring memory mappings after a TTM
+ * eviction.
+ *
+ * @invalidate_tlbs: Invalidate TLBs for a specific PASID
+ *
+ * @invalidate_tlbs_vmid: Invalidate TLBs for a specific VMID
+ *
  * This structure contains function pointers to services that the kgd driver
  * provides to amdkfd driver.
  *
@@ -275,6 +329,29 @@ struct kfd2kgd_calls {
 	void (*get_cu_info)(struct kgd_dev *kgd,
 			struct kfd_cu_info *cu_info);
 	uint64_t (*get_vram_usage)(struct kgd_dev *kgd);
+
+	int (*create_process_vm)(struct kgd_dev *kgd, void **vm,
+			void **process_info, struct dma_fence **ef);
+	void (*destroy_process_vm)(struct kgd_dev *kgd, void *vm);
+	uint32_t (*get_process_page_dir)(void *vm);
+	void (*set_vm_context_page_table_base)(struct kgd_dev *kgd,
+			uint32_t vmid, uint32_t page_table_base);
+	int (*alloc_memory_of_gpu)(struct kgd_dev *kgd, uint64_t va,
+			uint64_t size, void *vm,
+			struct kgd_mem **mem, uint64_t *offset,
+			uint32_t flags);
+	int (*free_memory_of_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem);
+	int (*map_memory_to_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem,
+			void *vm);
+	int (*unmap_memory_to_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem,
+			void *vm);
+	int (*sync_memory)(struct kgd_dev *kgd, struct kgd_mem *mem, bool intr);
+	int (*map_gtt_bo_to_kernel)(struct kgd_dev *kgd, struct kgd_mem *mem,
+			void **kptr, uint64_t *size);
+	int (*restore_process_bos)(void *process_info, struct dma_fence **ef);
+
+	int (*invalidate_tlbs)(struct kgd_dev *kgd, uint16_t pasid);
+	int (*invalidate_tlbs_vmid)(struct kgd_dev *kgd, uint16_t vmid);
 };
 
 /**