/*
* 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
dprint(FD_MEM, "pinning %llu bytes\n", td->o.lockmem);
/*
- * Don't allow mlock of more than real_mem-128MB
+ * Don't allow mlock of more than real_mem-128MiB
*/
phys_mem = os_phys_mem();
if (phys_mem) {
if ((td->o.lockmem + 128 * 1024 * 1024) > phys_mem) {
td->o.lockmem = phys_mem - 128 * 1024 * 1024;
- log_info("fio: limiting mlocked memory to %lluMB\n",
+ log_info("fio: limiting mlocked memory to %lluMiB\n",
td->o.lockmem >> 20);
}
}
}
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) {
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);
}
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);
}
}
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.
*/
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;
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);