Perftest not starting if --use_cuda_dmabuf passed before --use_cuda

[skondurian]

./ib_write_bw --version
Version: 6.25

./ib_write_bw -d mlx5_0 -S 0 --report_gbits -m 4096 -x 3 --qp 4 -a -F --tclass 224 -b --use_cuda_dmabuf --ipv6-addr --use_data_direct --use_cuda=1 -p 50001
CUDA DMA-BUF cannot be used without CUDA
 Parser function exited with Error

Same command with --use_cuda rearranged before --use_cuda_dmabuf

./ib_write_bw -d mlx5_0 -S 0 --report_gbits -m 4096 -x 3 --qp 4 -a -F --tclass 224 -b --use_cuda=1 --use_cuda_dmabuf --ipv6-addr --use_data_direct -p 50001
Perftest doesn't supports CUDA tests with inline messages: inline size set to 0

************************************
* Waiting for client to connect... *
************************************
^C

[Ren-hongchen]

Looking at the code, I found that perftest parses the incoming arguments one by one in a loop. When it encounters --use_cuda_dmabuf, it immediately validates that memory_type must be MEMORY_CUDA; this type is supposed to be set earlier by the --use_cuda argument. If --use_cuda_dmabuf is placed before --use_cuda, the validator runs as soon as --use_cuda_dmabuf is seen, but memory_type has not been assigned yet, so the program exits with an error.

if (use_cuda_flag) {
	CHECK_VALUE_NON_NEGATIVE(user_param->cuda_device_id,int,"CUDA device",not_int_ptr);
	user_param->memory_type = MEMORY_CUDA;   // assign MEMORY_CUDA in here
	user_param->memory_create = cuda_memory_create;
	use_cuda_flag = 0;
}
if (use_cuda_bus_id_flag) {
	user_param->cuda_device_bus_id = strdup(optarg);
	printf("Got PCIe address of: %s\n", user_param->cuda_device_bus_id);
	user_param->memory_type = MEMORY_CUDA;
	user_param->memory_create = cuda_memory_create;
	use_cuda_bus_id_flag = 0;
}
if (use_cuda_dmabuf_flag) { // if we parse use_cuda_dmabuf first, memory_type hasn't been assigned
	user_param->use_cuda_dmabuf = 1;
	if (user_param->memory_type != MEMORY_CUDA) {   // exit here
		fprintf(stderr, "CUDA DMA-BUF cannot be used without CUDA\n");
		free(duplicates_checker);
		return FAILURE;
	}
	use_cuda_dmabuf_flag = 0;
}

Several other arguments in the current code base exhibit the same issue. I'm therefore unsure whether this ordering requirement is intentional in perftest.
If we are allowed to change it, we could defer all validation logic until after the argument-parsing loop and only set flag bits during parsing.