hello，If I want to support GPUDirect Storage in the new storage system, in addition to the support of nvidia-fs.ko and new-fs-client.ko, do I need to do related adaptations in the cuda library?
It means that when nvidia-fs.ko supports the new file system, and the new-fs-client.ko also supports RDMA and GDS-related API calls, does it need to be supported at the cuda library level?
// These are same key values as the one’s in the user space.
// When adding a new driver make sure to use the same values in user and kernel space.
#define NVFS_PROC_MOD_NVME_KEY “nvme”
#define NVFS_PROC_MOD_DDN_LUSTRE_KEY “lnet”
#define NVFS_PROC_MOD_GPFS_KEY “mmfslinux”
#define NVFS_PROC_MOD_NFS_KEY “rpcrdma”
Take nfs as an example,
First need to increase symbol item in nvidia-fs.ko
struct module_entry modules_list = // nvfs symbol table
{
1,
0,
NVFS_PROC_MOD_NFS_KEY,
0,
“rpcrdma_register_nvfs_dma_ops”,
0,
“rpcrdma_unregister_nvfs_dma_ops”,
0,
&nvfs_dev_dma_rw_ops
},
Secondly, rpcrdma_register_nvfs_dma_ops is supported in the nfs driver.
In addition to the above two steps, do I need special settings or support for the cuda library?