Re: [PATCH bpf-next v2 00/28] Introduce eBPF support for HID devices

Hi Song,

Thanks a lot for the review.

I'll comment on the review in more details next week, but I have a
quick question here:

On Sat, Mar 5, 2022 at 2:14 AM Song Liu [off-list ref] wrote:

On Fri, Mar 4, 2022 at 9:29 AM Benjamin Tissoires
[off-list ref] wrote:

quoted

Hi,

This is a followup of my v1 at [0].

The short summary of the previous cover letter and discussions is that
HID could benefit from BPF for the following use cases:

- simple fixup of report descriptor:
  benefits are faster development time and testing, with the produced
  bpf program being shipped in the kernel directly (the shipping part
  is *not* addressed here).

- Universal Stylus Interface:
  allows a user-space program to define its own kernel interface

- Surface Dial:
  somehow similar to the previous one except that userspace can decide
  to change the shape of the exported device

- firewall:
  still partly missing there, there is not yet interception of hidraw
  calls, but it's coming in a followup series, I promise

- tracing:
  well, tracing.


I tried to address as many comments as I could and here is the short log
of changes:

v2:
===

- split the series by subsystem (bpf, HID, libbpf, selftests and
  samples)

- Added an extra patch at the beginning to not require CAP_NET_ADMIN for
  BPF_PROG_TYPE_LIRC_MODE2 (please shout if this is wrong)

- made the bpf context attached to HID program of dynamic size:
  * the first 1 kB will be able to be addressed directly
  * the rest can be retrieved through bpf_hid_{set|get}_data
    (note that I am definitivey not happy with that API, because there
    is part of it in bits and other in bytes. ouch)

- added an extra patch to prevent non GPL HID bpf programs to be loaded
  of type BPF_PROG_TYPE_HID
  * same here, not really happy but I don't know where to put that check
    in verifier.c

- added a new flag BPF_F_INSERT_HEAD for BPF_LINK_CREATE syscall when in
  used with HID program types.
  * this flag is used for tracing, to be able to load a program before
    any others that might already have been inserted and that might
    change the data stream.

Cheers,
Benjamin

The set looks good so far. I will review the rest later.

[...]

A quick note about how we organize these patches. Maybe we can
merge some of these patches like:

Just to be sure we are talking about the same thing: you mean squash
the patch together?

quoted

  bpf: introduce hid program type
  bpf/hid: add a new attach type to change the report descriptor
  bpf/hid: add new BPF type to trigger commands from userspace

I guess the three can merge into one.

quoted

  HID: hook up with bpf
  HID: allow to change the report descriptor from an eBPF program
  HID: bpf: compute only the required buffer size for the device
  HID: bpf: only call hid_bpf_raw_event() if a ctx is available

I haven't read through all of them, but I guess they can probably merge
as well.

There are certainly patches that we could squash together (3 and 4
from this list into the previous ones), but I'd like to keep some sort
of granularity here to not have a patch bomb that gets harder to come
back later.

quoted

  libbpf: add HID program type and API
  libbpf: add new attach type BPF_HID_RDESC_FIXUP
  libbpf: add new attach type BPF_HID_USER_EVENT

There 3 can merge, and maybe also the one below

quoted

  libbpf: add handling for BPF_F_INSERT_HEAD in HID programs

Yeah, the libbpf changes are small enough to not really justify
separate patches.

quoted

  samples/bpf: add new hid_mouse example
  samples/bpf: add a report descriptor fixup
  samples/bpf: fix bpf_program__attach_hid() api change

Maybe it makes sense to merge these 3?

Sure, why not.

quoted

  bpf/hid: add hid_{get|set}_data helpers
  HID: bpf: implement hid_bpf_get|set_data
  bpf/hid: add bpf_hid_raw_request helper function
  HID: add implementation of bpf_hid_raw_request

We can have 1 or 2 patches for these helpers

OK, the patches should be self-contained enough.

quoted

  selftests/bpf: add tests for the HID-bpf initial implementation
  selftests/bpf: add report descriptor fixup tests
  selftests/bpf: add tests for hid_{get|set}_data helpers
  selftests/bpf: add test for user call of HID bpf programs
  selftests/bpf: hid: rely on uhid event to know if a test device is
    ready
  selftests/bpf: add tests for bpf_hid_hw_request
  selftests/bpf: Add a test for BPF_F_INSERT_HEAD

These selftests could also merge into 1 or 2 patches I guess.

I'd still like to link them to the granularity of the bpf changes, so
I can refer a selftest change to a specific commit/functionality
added. But that's just my personal taste, and I can be convinced
otherwise. This should give us maybe 4 patches instead of 7.

I understand rearranging these patches may take quite some effort.
But I do feel that's a cleaner approach (from someone doesn't know
much about HID). If you really hate it that way, we can discuss...

No worries. I don't mind iterating on the series. IIRC I already
rewrote it twice from scratch, and that's when the selftests I
introduced in the second rewrite were tremendously helpful :) And
honestly I don't think it'll be too much effort to reorder/squash the
patches given that the v2 is *very* granular.

Anyway, I prefer having the reviewers happy so we can have a solid
rock API from day 1 than keeping it obscure for everyone and having to
deal with design issues forever. So if it takes 10 or 20 revisions to
have everybody on the same page, that's fine with me (not that I want
to have that many revisions, just that I won't be afraid of the
bikeshedding we might have at some point).

Cheers,
Benjamin