On Mon, May 02, 2022 at 11:43:51PM +0200, Benjamin Tissoires wrote:
> On Sat, Apr 30, 2022 at 9:12 AM Benjamin Tissoires
> <benjamin.tissoires@xxxxxxxxxx> wrote:
> >
> [...]
> > This is roughly what I have now:
> >
> > - hid-core is not aware of BPF except for a few __weak
> > ALLOW_ERROR_INJECTION hooks (dispatch_hid_bpf_device_event for
> > example)
> > - I have a separate hid-bpf module that attaches BPF traces to these
> > hooks and calls a "dispatch" kfunc within hid-bpf
> > - the dispatch function then do a succession of BPF calls to programs
> > attached to it by using bpf_tail_call(prog_array, hid_id)
> >
> > - for the clients, they define one or more
> > SEC("fmod_ret/hid_bpf_device_event"). That __weak hook is declared in
> > the kernel by hid-bpf but is never called, it's just an API
> > declaration
> > - then clients call in a SEC("syscall")
> > hid_bpf_attach_prog(ctx->prog_fd, ctx->hid_id, ctx->flags);
> > - hid_bpf_attach_prog is a kfunc that takes a ref on the struct
> > bpf_prog*, and stores that program in the correct struct bpf_map *for
> > the given attached_btf_id (hid_bpf_device_event in our case)
> >
> > And that's about it.
> > I still need to handle automatic release of the bpf prog when there is
> > no userspace open fd on it unless it's pinned but I think this should
> > be working fine.
> >
> > I also probably need to pin some SEC("syscall") (hid_bpf_attach_prog
> > and hid_bpf_dettach_prog) so users don't have to write them down and
> > can just use the ones provided by the kernel.
> >
> > The nice thing is that I can define my own API for the attach call
> > without dealing with bpf core. I can thus add a priority flag that is
> > relevant here because the data coming through the bpf program can be
> > modified.
> >
> > The other thing is that now, I don't care which function we are in to
> > decide if a RET_PTR_MEM is read only or not. I can deal with that by
> > either playing with the flags or even replacing entirely the dispatch
> > trace prog from userspace if I want to access the raw events.
> >
> > However, the downsides are:
> > - I need to also define kfuncs for BPF_PROG_TYPE_SYSCALL (I don't
> > think It'll be a big issue)
> > - The only way I could store the bpf_prog into the map was to hack
> > around the map ops, because the fd of the map in the skel is not
> > available while doing a SEC("syscall") from a different process.
>
> Update on this side: I realized that I could use the syscall
> BPF_MAP_GET_FD_BY_ID instead to get an fd for the current task.
> However, I've been bitten quite hard today because I was using
> bpf_map_get() instead of bpf_map_get_with_uref(), and so every time I
> closed the fd in the syscall the map was cleared...
>
> But now I would like to have more than one program of a type per hid
> device, meaning that I can not have only one bpf_map of type
> BPF_MAP_TYPE_PROG_ARRAY.
> I have explored BPF_MAP_TYPE_HASH_OF_MAPS, but we can not have
> BPF_MAP_TYPE_PROG_ARRAY as inner maps with the current code. And I'd
> need 2 levels of nesting (which is not authorized today):
> - hid_jmp_table (key: HID id)
> - array of different program type per HID device (key: HID_BPF_PROG_TYPE)
> - BPF_MAP_TYPE_PROG_ARRAY with the actual progs (key: int)
>
> The other solution would be to be able to create a map when needed,
> store it in struct hid_device, and then call bpf_tail_call on this
> map. The problem is that I need a way to teach the verifier that the
> struct bpf_map pointer I have in the context is a true bpf_map...
We have kptr feature now.
So bpf progs can store pointers to specific kernel data structures
inside map values.
Storing 'struct bpf_map *' in a map value would be something :)
Circular dependency issues to address. Maybe it's doable.
Would hash based prog_array work ?
Then the key can be an arbitrary combination.
There is fd_htab logic. It's used for map-in-map.
We can tweak it to store progs in a hash map.
