GPUDirect RDMA support

[fio.git] / memory.c

diff --git a/memory.c b/memory.c
index af4d5ef8ee371701f02806369a3581ecfdc42628..fe657225a2739bf878b2ddfb0909771ccb6e21ad 100644 (file)
--- a/memory.c
+++ b/memory.c
@@ -33,13 +33,13 @@ int fio_pin_memory(struct thread_data *td)
        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);
                }
        }
@@ -207,6 +207,75 @@ static void free_mem_malloc(struct thread_data *td)
        free(td->orig_buffer);
 }
 
+#ifdef CONFIG_CUDA
+
+static int alloc_mem_cudamalloc(struct thread_data *td, size_t total_mem)
+{
+       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;
+}
+
+static void free_mem_cudamalloc(struct thread_data *td)
+{
+       if ((void *) td->dev_mem_ptr != NULL)
+               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.
  */
@@ -215,13 +284,13 @@ int allocate_io_mem(struct thread_data *td)
        size_t total_mem;
        int ret = 0;
 
-       if (td->io_ops->flags & FIO_NOIO)
+       if (td_ioengine_flagged(td, FIO_NOIO))
                return 0;
 
        total_mem = td->orig_buffer_size;
 
        if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
-           (td->io_ops->flags & FIO_MEMALIGN)) {
+           td_ioengine_flagged(td, FIO_MEMALIGN)) {
                total_mem += page_mask;
                if (td->o.mem_align && td->o.mem_align > page_size)
                        total_mem += td->o.mem_align - page_size;
@@ -246,6 +315,10 @@ int allocate_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)
                ret = alloc_mem_mmap(td, total_mem);
+#ifdef CONFIG_CUDA
+       else if (td->o.mem_type == MEM_CUDA_MALLOC)
+               ret = alloc_mem_cudamalloc(td, total_mem);
+#endif
        else {
                log_err("fio: bad mem type: %d\n", td->o.mem_type);
                ret = 1;
@@ -275,6 +348,10 @@ 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);
+#ifdef CONFIG_CUDA
+       else if (td->o.mem_type == MEM_CUDA_MALLOC)
+               free_mem_cudamalloc(td);
+#endif
        else
                log_err("Bad memory type %u\n", td->o.mem_type);