MemoryRegion<T> is completely unsound as a concept #58
Description
The current MemoryRegion<T> allows you to store any arbitrary data type T inside a region of memory registered with libibverbs.
This allows two machines to copy the raw byte contents of that type T between them.
This is completely unsound and unsafe. The data type T could contain pointers to memory that is not registered with libibverbs:
struct MyType {
data: Vec<u64>, // THIS DATA LIVES ON THE HEAP, IS NOT MANAGED BY RDMA LIBRARY
}
// Try to receive one instance of MyType from a remote machine.
let mut mr: MemoryRegion<MyType> = pd.allocate::<MyType>(1).unwrap();
unsafe { qp.post_receive(&mut mr, 0, 1) }.unwrap();
// This is completely unsafe!! We read some arbitrary pointer from the remote machine's address space.
// That pointer doesn't point to anything reasonable in the local machine's address space.
// Not to mention, the remote machine could have a different endianness than the local machine.
println!("{:?}", mr[0].data);The only safe interface that you can expose for MemoryRegion is MemoryRegion<u8>. Technically, you could allow any type T that is "plain old data", but then you still have to worry about endianness, and it's just not worth it.
The only safe interface for RDMA is bytes.