Add hsaKmtRegisterMemoryWithFlags.

API follows hsaKmtRegisterMemory but allows passing HsaMemFlags.

Change-Id: I66a230a87c8b085f27c769bdf2cb4d0d96a5d6dd


[ROCm/ROCR-Runtime commit: c7f1277013]

projects/rocr-runtime/include/hsakmt.h

@@ -400,6 +400,18 @@ hsaKmtRegisterMemoryToNodes(
     );
 
 
+/**
+  Registers with KFD a memory buffer with memory attributes
+*/
+
+HSAKMT_STATUS
+HSAKMTAPI
+hsaKmtRegisterMemoryWithFlags(
+    void        *MemoryAddress,     // IN (cache-aligned)
+    HSAuint64   MemorySizeInBytes,  // IN (cache-aligned)
+    HsaMemFlags MemFlags            // IN
+    );
+
 /**
   Registers with KFD a graphics buffer and returns graphics metadata
 */
@@ -515,7 +527,7 @@ hsaKmtMapMemoryToGPUNodes(
     void*           MemoryAddress,         //IN (page-aligned)
     HSAuint64       MemorySizeInBytes,     //IN (page-aligned)
     HSAuint64*      AlternateVAGPU,        //OUT (page-aligned)
-     HsaMemMapFlags    MemMapFlags,               //IN
+    HsaMemMapFlags  MemMapFlags,           //IN
     HSAuint64       NumberOfNodes,         //IN
     HSAuint32*      NodeArray              //IN
     );

projects/rocr-runtime/src/fmm.c

@@ -2794,7 +2794,8 @@ bool fmm_get_handle(void *address, uint64_t *handle)
 	return found;
 }
 
-static HSAKMT_STATUS fmm_register_user_memory(void *addr, HSAuint64 size, vm_object_t **obj_ret)
+static HSAKMT_STATUS fmm_register_user_memory(void *addr, HSAuint64 size,
+				  vm_object_t **obj_ret, bool coarse_grain)
 {
 	manageable_aperture_t *aperture = svm.dgpu_aperture;
 	HSAuint32 page_offset = (HSAuint64)addr & (PAGE_SIZE-1);
@@ -2818,7 +2819,8 @@ static HSAKMT_STATUS fmm_register_user_memory(void *addr, HSAuint64 size, vm_obj
 	svm_addr = __fmm_allocate_device(gpu_id, NULL, aligned_size, aperture,
 			 &aligned_addr, KFD_IOC_ALLOC_MEM_FLAGS_USERPTR |
 			 KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE |
-			 KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE, &obj);
+			 KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE |
+			 (coarse_grain ? 0 : KFD_IOC_ALLOC_MEM_FLAGS_COHERENT), &obj);
 	if (!svm_addr)
 		return HSAKMT_STATUS_ERROR;
 
@@ -2841,7 +2843,8 @@ static HSAKMT_STATUS fmm_register_user_memory(void *addr, HSAuint64 size, vm_obj
 
 HSAKMT_STATUS fmm_register_memory(void *address, uint64_t size_in_bytes,
 				  uint32_t *gpu_id_array,
-				  uint32_t gpu_id_array_size)
+				  uint32_t gpu_id_array_size,
+				  bool coarse_grain)
 {
 	manageable_aperture_t *aperture = NULL;
 	vm_object_t *object = NULL;
@@ -2857,7 +2860,7 @@ HSAKMT_STATUS fmm_register_memory(void *address, uint64_t size_in_bytes,
 			return HSAKMT_STATUS_SUCCESS;
 
 		/* Register a new user ptr */
-		ret = fmm_register_user_memory(address, size_in_bytes, &object);
+		ret = fmm_register_user_memory(address, size_in_bytes, &object, coarse_grain);
 		if (ret != HSAKMT_STATUS_SUCCESS)
 			return ret;
 		if (gpu_id_array_size == 0)

projects/rocr-runtime/src/fmm.h

@@ -69,7 +69,8 @@ HSAKMT_STATUS fmm_get_aperture_base_and_limit(aperture_type_e aperture_type, HSA
 
 HSAKMT_STATUS fmm_register_memory(void *address, uint64_t size_in_bytes,
 								  uint32_t *gpu_id_array,
-								  uint32_t gpu_id_array_size);
+								  uint32_t gpu_id_array_size,
+								  bool coarse_grain);
 HSAKMT_STATUS fmm_register_graphics_handle(HSAuint64 GraphicsResourceHandle,
 					   HsaGraphicsResourceInfo *GraphicsResourceInfo,
 					   uint32_t *gpu_id_array,

projects/rocr-runtime/src/libhsakmt.ver

@@ -26,6 +26,7 @@ hsaKmtAllocMemory;
 hsaKmtFreeMemory;
 hsaKmtRegisterMemory;
 hsaKmtRegisterMemoryToNodes;
+hsaKmtRegisterMemoryWithFlags;
 hsaKmtRegisterGraphicsHandleToNodes;
 hsaKmtShareMemory;
 hsaKmtRegisterSharedHandle;

projects/rocr-runtime/src/memory.c

@@ -210,7 +210,7 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtRegisterMemory(void *MemoryAddress,
 		return HSAKMT_STATUS_SUCCESS;
 
 	return fmm_register_memory(MemoryAddress, MemorySizeInBytes,
-				   NULL, 0);
+				   NULL, 0, true);
 }
 
 HSAKMT_STATUS HSAKMTAPI hsaKmtRegisterMemoryToNodes(void *MemoryAddress,
@@ -235,7 +235,8 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtRegisterMemoryToNodes(void *MemoryAddress,
 	if (ret == HSAKMT_STATUS_SUCCESS) {
 		ret = fmm_register_memory(MemoryAddress, MemorySizeInBytes,
 					  gpu_id_array,
-					  NumberOfNodes*sizeof(uint32_t));
+					  NumberOfNodes*sizeof(uint32_t),
+					  true);
 		if (ret != HSAKMT_STATUS_SUCCESS)
 			free(gpu_id_array);
 	}
@@ -243,6 +244,30 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtRegisterMemoryToNodes(void *MemoryAddress,
 	return ret;
 }
 
+HSAKMT_STATUS HSAKMTAPI hsaKmtRegisterMemoryWithFlags(void *MemoryAddress,
+						    HSAuint64 MemorySizeInBytes,
+						    HsaMemFlags MemFlags)
+{
+	CHECK_KFD_OPEN();
+	HSAKMT_STATUS ret = HSAKMT_STATUS_SUCCESS;
+
+	pr_debug("[%s] address %p\n",
+		__func__, MemoryAddress);
+
+	// Registered memory should be ordinary paged host memory.
+	if ((MemFlags.ui32.HostAccess != 1) || (MemFlags.ui32.NonPaged == 1))
+		return HSAKMT_STATUS_NOT_SUPPORTED;
+
+	if (!is_dgpu)
+		/* TODO: support mixed APU and dGPU configurations */
+		return HSAKMT_STATUS_NOT_SUPPORTED;
+
+	ret = fmm_register_memory(MemoryAddress, MemorySizeInBytes,
+		NULL, 0, MemFlags.ui32.CoarseGrain);
+
+	return ret;
+}
+
 HSAKMT_STATUS HSAKMTAPI hsaKmtRegisterGraphicsHandleToNodes(HSAuint64 GraphicsResourceHandle,
 							    HsaGraphicsResourceInfo *GraphicsResourceInfo,
 							    HSAuint64 NumberOfNodes,