X-Git-Url: https://git.kernel.dk/?a=blobdiff_plain;f=memory.c;h=577d3dd5afd80d2a3dd1cdd4183b797b276f0f04;hb=615c794cbf851c994e94fffe8b8f565e64f137a5;hp=9e73f100740a20c0f7d5c17af23366d92e436186;hpb=420b104a3f6ba4328b6b167808f71bb7849102ec;p=fio.git

diff --git a/memory.c b/memory.c
index 9e73f100..577d3dd5 100644
--- a/memory.c
+++ b/memory.c
@@ -1,11 +1,10 @@
 /*
  * Memory helpers
  */
-#include <sys/types.h>
-#include <sys/stat.h>
 #include <fcntl.h>
 #include <unistd.h>
 #include <sys/mman.h>
+#include <sys/stat.h>
 
 #include "fio.h"
 #ifndef FIO_NO_HAVE_SHM_H
@@ -138,6 +137,9 @@ static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
 	}
 
 	if (td->o.mmapfile) {
+		if (access(td->o.mmapfile, F_OK) == 0)
+			td->flags |= TD_F_MMAP_KEEP;
+
 		td->mmapfd = open(td->o.mmapfile, O_RDWR|O_CREAT, 0644);
 
 		if (td->mmapfd < 0) {
@@ -169,7 +171,7 @@ static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
 		td->orig_buffer = NULL;
 		if (td->mmapfd != 1 && td->mmapfd != -1) {
 			close(td->mmapfd);
-			if (td->o.mmapfile)
+			if (td->o.mmapfile && !(td->flags & TD_F_MMAP_KEEP))
 				unlink(td->o.mmapfile);
 		}
 
@@ -187,7 +189,8 @@ static void free_mem_mmap(struct thread_data *td, size_t total_mem)
 	if (td->o.mmapfile) {
 		if (td->mmapfd != -1)
 			close(td->mmapfd);
-		unlink(td->o.mmapfile);
+		if (!(td->flags & TD_F_MMAP_KEEP))
+			unlink(td->o.mmapfile);
 		free(td->o.mmapfile);
 	}
 }
@@ -207,6 +210,78 @@ static void free_mem_malloc(struct thread_data *td)
 	free(td->orig_buffer);
 }
 
+static int alloc_mem_cudamalloc(struct thread_data *td, size_t total_mem)
+{
+#ifdef CONFIG_CUDA
+	CUresult ret;
+	char name[128];
+
+	ret = cuInit(0);
+	if (ret != CUDA_SUCCESS) {
+		log_err("fio: failed initialize cuda driver api\n");
+		return 1;
+	}
+
+	ret = cuDeviceGetCount(&td->gpu_dev_cnt);
+	if (ret != CUDA_SUCCESS) {
+		log_err("fio: failed get device count\n");
+		return 1;
+	}
+	dprint(FD_MEM, "found %d GPU devices\n", td->gpu_dev_cnt);
+
+	if (td->gpu_dev_cnt == 0) {
+		log_err("fio: no GPU device found. "
+			"Can not perform GPUDirect RDMA.\n");
+		return 1;
+	}
+
+	td->gpu_dev_id = td->o.gpu_dev_id;
+	ret = cuDeviceGet(&td->cu_dev, td->gpu_dev_id);
+	if (ret != CUDA_SUCCESS) {
+		log_err("fio: failed get GPU device\n");
+		return 1;
+	}
+
+	ret = cuDeviceGetName(name, sizeof(name), td->gpu_dev_id);
+	if (ret != CUDA_SUCCESS) {
+		log_err("fio: failed get device name\n");
+		return 1;
+	}
+	dprint(FD_MEM, "dev_id = [%d], device name = [%s]\n", \
+	       td->gpu_dev_id, name);
+
+	ret = cuCtxCreate(&td->cu_ctx, CU_CTX_MAP_HOST, td->cu_dev);
+	if (ret != CUDA_SUCCESS) {
+		log_err("fio: failed to create cuda context: %d\n", ret);
+		return 1;
+	}
+
+	ret = cuMemAlloc(&td->dev_mem_ptr, total_mem);
+	if (ret != CUDA_SUCCESS) {
+		log_err("fio: cuMemAlloc %zu bytes failed\n", total_mem);
+		return 1;
+	}
+	td->orig_buffer = (void *) td->dev_mem_ptr;
+
+	dprint(FD_MEM, "cudaMalloc %llu %p\n",				\
+	       (unsigned long long) total_mem, td->orig_buffer);
+	return 0;
+#else
+	return -EINVAL;
+#endif
+}
+
+static void free_mem_cudamalloc(struct thread_data *td)
+{
+#ifdef CONFIG_CUDA
+	if (td->dev_mem_ptr)
+		cuMemFree(td->dev_mem_ptr);
+
+	if (cuCtxDestroy(td->cu_ctx) != CUDA_SUCCESS)
+		log_err("fio: failed to destroy cuda context\n");
+#endif
+}
+
 /*
  * Set up the buffer area we need for io.
  */
@@ -230,22 +305,26 @@ int allocate_io_mem(struct thread_data *td)
 	dprint(FD_MEM, "Alloc %llu for buffers\n", (unsigned long long) total_mem);
 
 	/*
-	 * If the IO engine has hooks to allocate/free memory, use those. But
-	 * error out if the user explicitly asked for something else.
+	 * If the IO engine has hooks to allocate/free memory and the user
+	 * doesn't explicitly ask for something else, use those. But fail if the
+	 * user asks for something else with an engine that doesn't allow that.
 	 */
-	if (td->io_ops->iomem_alloc) {
-		if (fio_option_is_set(&td->o, mem_type)) {
-			log_err("fio: option 'mem/iomem' conflicts with specified IO engine\n");
-			ret = 1;
-		} else
-			ret = td->io_ops->iomem_alloc(td, total_mem);
-	} else if (td->o.mem_type == MEM_MALLOC)
+	if (td->io_ops->iomem_alloc && fio_option_is_set(&td->o, mem_type) &&
+	    !td_ioengine_flagged(td, FIO_SKIPPABLE_IOMEM_ALLOC)) {
+		log_err("fio: option 'mem/iomem' conflicts with specified IO engine\n");
+		ret = 1;
+	} else if (td->io_ops->iomem_alloc &&
+		   !fio_option_is_set(&td->o, mem_type))
+		ret = td->io_ops->iomem_alloc(td, total_mem);
+	else if (td->o.mem_type == MEM_MALLOC)
 		ret = alloc_mem_malloc(td, total_mem);
 	else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE)
 		ret = alloc_mem_shm(td, total_mem);
 	else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE ||
 		 td->o.mem_type == MEM_MMAPSHARED)
 		ret = alloc_mem_mmap(td, total_mem);
+	else if (td->o.mem_type == MEM_CUDA_MALLOC)
+		ret = alloc_mem_cudamalloc(td, total_mem);
 	else {
 		log_err("fio: bad mem type: %d\n", td->o.mem_type);
 		ret = 1;
@@ -265,7 +344,7 @@ void free_io_mem(struct thread_data *td)
 	if (td->o.odirect || td->o.oatomic)
 		total_mem += page_mask;
 
-	if (td->io_ops->iomem_alloc) {
+	if (td->io_ops->iomem_alloc && !fio_option_is_set(&td->o, mem_type)) {
 		if (td->io_ops->iomem_free)
 			td->io_ops->iomem_free(td);
 	} else if (td->o.mem_type == MEM_MALLOC)
@@ -275,6 +354,8 @@ void free_io_mem(struct thread_data *td)
 	else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE ||
 		 td->o.mem_type == MEM_MMAPSHARED)
 		free_mem_mmap(td, total_mem);
+	else if (td->o.mem_type == MEM_CUDA_MALLOC)
+		free_mem_cudamalloc(td);
 	else
 		log_err("Bad memory type %u\n", td->o.mem_type);