On Friday 2020-04-24 14:15, Jason Gunthorpe wrote: >> ./usr/include/rdma/ib_user_verbs.h:436:34: warning: field 'base' with >> variable sized type 'struct ib_uverbs_create_cq_resp' not at the end of >> a struct or class is a GNU extension >> [-Wgnu-variable-sized-type-not-at-end] >> struct ib_uverbs_create_cq_resp base; >> ^ >> I presume this is part of the point of the conversion since you mention >> a compiler warning when the flexible member is not at the end of a >> struct. How should they be fixed? That should probably happen before the >> patch gets merged. > >The flexible member IS at the end of the struct and is often intended >to cover the memory in the enclosing struct. There is no guarantee for the presence of such a struct. In the case of the RDMA headers, even if we assume best conditions -- a block of malloc(sizeof(struct ib_uverbs_create_cq_resp) + sizeof(u64)*N) and not some struct -- it smells a lot like undefined behavior, because ib_uverbs_create_cq_resp::driver_data accesses data as u64 while ib_uverbs_ex_create_cq_resp::comp_mask and friends are u32. There has got to be some aliasing rule in C that causes RDMA's purported use-case to be UB.
