Re: [PATCH bpf-next v3 02/17] bpf: introduce hid program type

[Benjamin Tissoires]

On Mon, Mar 21, 2022 at 10:52 PM Song Liu <song@xxxxxxxxxx> wrote:
>
> On Mon, Mar 21, 2022 at 9:07 AM Benjamin Tissoires
> <benjamin.tissoires@xxxxxxxxxx> wrote:
> >
> > Hi Song,
> >
> > many thanks for the quick response.
> >
> > On Fri, Mar 18, 2022 at 9:48 PM Song Liu <song@xxxxxxxxxx> wrote:
> [...]
> > >
> > > > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> > >
> > > We need to mirror these changes to tools/include/uapi/linux/bpf.h.
> >
> > OK. I did that in patch 4/17 but I can bring in the changes there too.
>
> Let's keep changes to the two files in the same patch. This will make
> sure they are back ported together.

Ack

>
> [...]
> > > > +enum hid_bpf_event {
> > > > +       HID_BPF_UNDEF = 0,
> > > > +       HID_BPF_DEVICE_EVENT,           /* when attach type is BPF_HID_DEVICE_EVENT */
> > > > +       HID_BPF_RDESC_FIXUP,            /* ................... BPF_HID_RDESC_FIXUP */
> > > > +       HID_BPF_USER_EVENT,             /* ................... BPF_HID_USER_EVENT */
> > >
> > > Why don't we have a DRIVER_EVENT type here?
> >
> > For driver event, I want to have a little bit more of information
> > which tells which event we have:
> > - HID_BPF_DRIVER_PROBE
> > - HID_BPF_DRIVER_SUSPEND
> > - HID_BPF_DRIVER_RAW_REQUEST
> > - HID_BPF_DRIVER_RAW_REQUEST_ANSWER
> > - etc...
> >
> > However, I am not entirely sure on the implementation of all of those,
> > so I left them aside for now.
> >
> > I'll work on that for v4.
>
> This set is already pretty big. I guess we can add them in a follow-up set.
>
> >
> > >
> > > >
> > > [...]
 [...]
> > > > +
> > > > +static int hid_bpf_prog_test_run(struct bpf_prog *prog,
> > > > +                                const union bpf_attr *attr,
> > > > +                                union bpf_attr __user *uattr)
> > > > +{
> > > > +       struct hid_device *hdev = NULL;
> > > > +       struct bpf_prog_array *progs;
> > > > +       bool valid_prog = false;
> > > > +       int i;
> > > > +       int target_fd, ret;
> > > > +       void __user *data_out = u64_to_user_ptr(attr->test.data_out);
> > > > +       void __user *data_in = u64_to_user_ptr(attr->test.data_in);
> > > > +       u32 user_size_in = attr->test.data_size_in;
> > > > +       u32 user_size_out = attr->test.data_size_out;
> > > > +       u32 allocated_size = max(user_size_in, user_size_out);
> > > > +       struct hid_bpf_ctx_kern ctx = {
> > > > +               .type = HID_BPF_USER_EVENT,
> > > > +               .allocated_size = allocated_size,
> > > > +       };
> > > > +
> > > > +       if (!hid_hooks.hdev_from_fd)
> > > > +               return -EOPNOTSUPP;
> > > > +
> > > > +       if (attr->test.ctx_size_in != sizeof(int))
> > > > +               return -EINVAL;
> > >
> > > ctx_size_in is always 4 bytes?
> >
> > Yes. Basically what I had in mind is that the "ctx" for
> > user_prog_test_run is the file descriptor to the sysfs that represent
> > the HID device.
> > This seemed to me to be the easiest to handle for users.
> >
> > I'm open to suggestions though.
>
> How about we use data_in? ctx for test_run usually means the program ctx,
> which is struct hid_bpf_ctx here.
>

I'd rather not use data_in. data_in is forwarded as it is in the ctx
of the program, so adding a bulky API where the first byte is the
target_fd doesn't make a lot of sense IMO.

However, I just managed to achieve what I initially wanted to do
without luck: just use the struct bpf_prog as the sole argument.
I thought iterating over all hid devices would be painful, but it
turns out that is exactly what hid_bpf_fd_to_hdev() was doing, so
there is no penalty in doing so.

Anyway, I'll drop ctx_in in the next version.

Cheers,
Benjamin