On Mon, Sep 23, 2019 at 09:15:54AM +0000, Michal Kalderon wrote: > > This hidden detail should also be noted in the comment for > > rdma_user_mmap_io.. > > I actually misunderstood you before and thought that we want to > maintain all mappings in umap. Now I understand you meant only the > io mappings, since they're not referenced counted by the mm system. > I'll revert the changes and add an additional parameter to > rdma_user_mmap_io -> However, this means more changes in drivers > that call this function and don't use the mmap_xa, Should I add an > additional interface ? or are you OK with me sending NULL to the > additional parameter from all drivers using the interface ? (hns, > mlx4, mlx5) It should be OK > Also, who should increase the refcnt of rdma_user_mmap_entry when it > is set to the vma private data ? The caller of rdma_user_mmap_io or > inside rdma_user_mmap_io function ? This will definitely simplify > things. I generally prefer to see refcounts incrd at the spot the pointer is copied, so the thing that sets vm_prviate_data should incr the refcount. Jason
