Re: [PATCH RFC bpf-next 1/3] bpf: revamp bpf_attr and name each command's field and substruct

[Andrii Nakryiko <andrii.nakryiko@gmail.com>]

On Tue, May 30, 2023 at 10:41 AM Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
>
> On Thu, May 25, 2023 at 4:40 PM Andrii Nakryiko
> <andrii.nakryiko@gmail.com> wrote:
> >
> > On Thu, May 25, 2023 at 2:51 PM Daniel Borkmann <daniel@iogearbox.net> wrote:
> > >
> > > On 5/25/23 7:19 PM, Andrii Nakryiko wrote:
> > > > On Wed, May 24, 2023 at 8:18 PM Alexei Starovoitov
> > > > <alexei.starovoitov@gmail.com> wrote:
> > > >> On Wed, May 24, 2023 at 02:02:41PM -0700, Andrii Nakryiko wrote:
> > > >>>
> > > >>> And there were a bunch of other similar changes. Please take a thorough
> > > >>> look and suggest more changes or which changes to drop. I'm not married
> > > >>> to any of them, it just felt like a good improvement.
> > > >>
> > > >> Agree that current layout sucks, but ...
> > > >>
> > > >>>   include/uapi/linux/bpf.h       | 235 +++++++++++++++++++++++++++------
> > > >>>   kernel/bpf/syscall.c           |  40 +++---
> > > >>>   tools/include/uapi/linux/bpf.h | 235 +++++++++++++++++++++++++++------
> > > >>>   3 files changed, 405 insertions(+), 105 deletions(-)
> > > >>
> > > >> ... the diff makes it worse. The diffstat for "nop" change is a red flag.
> > > >
> > > > Only 100 lines are a real "nop" change to copy/paste existing fields
> > > > that are in unnamed fields. The rest is a value add.
> > > >
> > > > I don't think the deal is in stats, though, right?
> > > >
> > > >>> +     /*
> > > >>> +      * LEGACY anonymous substructs, for backwards compatibility.
> > > >>> +      * Each of the below anonymous substructs are ABI compatible with one
> > > >>> +      * of the above named substructs. Please use named substructs.
> > > >>> +      */
> > > >>
> > > >> All of them cannot be removed. This bagage will be a forever eyesore.
> > > >> Currently it's not pretty. The diffs make uapi file just ugly.
> > > >> Especially considering how 'named' and 'legacy' will start diverging.
> > > >
> > > > We have to allow "divergence" (only in the sense that new fields only
> > > > go into named substructs, but the existing fields stay fixed, of
> > > > course), to avoid more naming conflicts. If that wasn't the case,
> > > > using struct_group() macro could have been used to avoid a copy/paste
> > > > of those anonymous field/struct copies.
> > > >
> > > > So I'm not happy about those 100 lines copy paste of fixed fields
> > > > either, but at least that would get us out of the current global
> > > > naming namespace for PROG_LOAD, MAP_CREATE, BTF_LOAD, etc.
> > > >
> > > >> New commands are thankfully named. We've learned the lesson,
> > > >
> > > > Unfortunately, the problem is that unnamed commands are the ones that
> > > > are most likely to keep evolving.
> > > >
> > > >> but prior mistake is unfixable. We have to live with it.
> > > >
> > > > Ok, too bad, but it's fine. It was worth a try.
> > > >
> > > > I tried to come up with something like struct_group() approach to
> > > > minimize code changes in UAPI header, but we have a more complicated
> > > > situation where part of struct has to be both anonymous and named,
> > > > while another part (newly added fields) should go only to named parts.
> > > > And that doesn't seem to be possible to support with a macro,
> > > > unfortunately.
> > >
> > > Nice idea on the struct_group()-like approach, but agree that this is
> > > going to be tough given we need to divert anonymous and named parts as
> > > you mention. One other wild thought ... we remove the bpf_attr entirely
> > > from the uapi header, and have a kernel/bpf/bpf.cmd description and
> > > generate the bpf_attr into a uapi header via script which the main header
> > > can include. Kind of similar to the suggestion, but more flexible than
> > > macro magic. We also have things like syscall table header generated via
> > > script.. so it wouldn't be first. Still doesn't remove the eyesore, just
> > > packages it differently. ;/
> >
> > There are two more ways, neither is that pretty. But I'll just outline
> > them here for completeness.
> >
> > First, we can define about 6 variants (one for each command with anon
> > field) of macro with different numbers of arguments, one for each
> > existing field. Replace all semicolons with commas and do something
> > like this (we can prettify the below some more, I didn't want to waste
> > too much time on this demo):
> >
> > #define __bpf_cmd4(type, f1, f2, f3, f4, new_fields...)        \
> >        struct {                                                \
> >                f1; f2; f3; f4;                                 \
> >        };                                                      \
> >        struct type {                                           \
> >                f1; f2; f3; f4;                                 \
> >                new_fields                                      \
> >        }
> >
> >        /* BPF_OBJ_PIN command */
> >        __bpf_cmd4(bpf_obj_pin_attr,
> >                __aligned_u64   pathname,
> >                __u32           bpf_fd,
> >                __u32           file_flags,
> >                /* Same as dirfd in openat() syscall; see openat(2)
> >                 * manpage for details of path FD and pathname semantics;
> >                 * path_fd should accompanied by BPF_F_PATH_FD flag set in
> >                 * file_flags field, otherwise it should be set to zero;
> >                 * if BPF_F_PATH_FD flag is not set, AT_FDCWD is assumed.
> >                 */
> >                __s32           path_fd,
> >                __u32           token_fd;
> >        ) obj_pin;
> >
> > Note that I also added `__u32 token_fd;` as a demonstration how we can
> > new fields, and that new fields will have proper semicolons at the
> > end. The largest command (BPF_PROG_LOAD) will need 28 arg variant, but
> > that can be fit in few lines pretty cleanly, if the overall approach
> > would be deemed acceptable.
> >
> > This approach also has a slight downside that we can rename fields
> > (e.g. for BPF_BTF_LOAD command). We still can split out dedicated new
> > named structs. So too big of a deal.
> >
> >
> > Second approach. If it's mostly about aesthetics, then we can add
> > include/uapi/linux/bpf_legacy.h, where we put all these unnamed fields
> > and structs in one stashed away place, and then in original
> > include/uapi/linux/bpf.h header we just
> >
> > union bpf_attr {
> >    ... named structs and fields go here ...
> >
> > /* include backwards compat legacy anon fields/structs */
> > #include "bpf_legacy.h"
> > };
> >
> > This way this eyesore will be somewhat hidden away (but still lookup-able).
> >
> >
> > Curious if any of the above is more palatable?
>
> Frankly I don't like either Daniel's .cmd idea or these two "aesthetics".
> We just need new *_token_fd fields in several structures.
> imo adding several such fields with different prefixes are cleaner
> than revamping the whole thing.

Ok, sgtm.