On Mon, Sep 20, 2021 at 7:15 AM Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx> wrote: > > This change updates the BPF syscall loader to relocate BTF_KIND_FUNC > relocations, with support for weak kfunc relocations. The next commit > adds bpftool supports to set up the fd_array_sz parameter for light > skeleton. > > A second map for keeping fds is used instead of adding fds to existing > loader.map because of following reasons: but it complicates signing bpf progs a lot. > If reserving an area for map and BTF fds, we would waste the remaining > of (MAX_USED_MAPS + MAX_KFUNC_DESCS) * sizeof(int), which in most cases > will be unused by the program. Also, we must place some limit on the > amount of map and BTF fds a program can possibly open. That is just (256 + 64)*4 bytes of data. Really not much. I wouldn't worry about reserving this space. > If setting gen->fd_array to first map_fd offset, and then just finding > the offset relative to this (for later BTF fds), such that they can be > packed without wasting space, we run the risk of unnecessarily running > out of valid offset for emit_relo stage (for kfuncs), because gen map > creation and relocation stages are separated by other steps that can add > lots of data (including bpf_object__populate_internal_map). It is also > prone to break silently if features are added between map and BTF fd > emits that possibly add more data (just ~128KB to break BTF fd, since > insn->off allows for INT16_MAX (32767) * 4 bytes). I don't follow this logic. > Both of these issues are compounded by the fact that data map is shared > by all programs, so it is easy to end up with invalid offset for BTF fd. I don't follow this either. There is only one map and one program. What sharing are you talking about?
