Re: [PATCH v2 bpf-next 05/14] bpf: allow to specify user-provided context value for BPF perf links

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

On Thu, Jul 29, 2021 at 10:49 PM Yonghong Song <yhs@fb.com> wrote:
>
>
>
> On 7/29/21 9:31 PM, Andrii Nakryiko wrote:
> > On Thu, Jul 29, 2021 at 11:00 AM Yonghong Song <yhs@fb.com> wrote:
> >>
> >>
> >>
> >> On 7/26/21 9:12 AM, Andrii Nakryiko wrote:
> >>> Add ability for users to specify custom u64 value when creating BPF link for
> >>> perf_event-backed BPF programs (kprobe/uprobe, perf_event, tracepoints).
> >>>
> >>> This is useful for cases when the same BPF program is used for attaching and
> >>> processing invocation of different tracepoints/kprobes/uprobes in a generic
> >>> fashion, but such that each invocation is distinguished from each other (e.g.,
> >>> BPF program can look up additional information associated with a specific
> >>> kernel function without having to rely on function IP lookups). This enables
> >>> new use cases to be implemented simply and efficiently that previously were
> >>> possible only through code generation (and thus multiple instances of almost
> >>> identical BPF program) or compilation at runtime (BCC-style) on target hosts
> >>> (even more expensive resource-wise). For uprobes it is not even possible in
> >>> some cases to know function IP before hand (e.g., when attaching to shared
> >>> library without PID filtering, in which case base load address is not known
> >>> for a library).
> >>>
> >>> This is done by storing u64 user_ctx in struct bpf_prog_array_item,
> >>> corresponding to each attached and run BPF program. Given cgroup BPF programs
> >>> already use 2 8-byte pointers for their needs and cgroup BPF programs don't
> >>> have (yet?) support for user_ctx, reuse that space through union of
> >>> cgroup_storage and new user_ctx field.
> >>>
> >>> Make it available to kprobe/tracepoint BPF programs through bpf_trace_run_ctx.
> >>> This is set by BPF_PROG_RUN_ARRAY, used by kprobe/uprobe/tracepoint BPF
> >>> program execution code, which luckily is now also split from
> >>> BPF_PROG_RUN_ARRAY_CG. This run context will be utilized by a new BPF helper
> >>> giving access to this user context value from inside a BPF program. Generic
> >>> perf_event BPF programs will access this value from perf_event itself through
> >>> passed in BPF program context.
> >>>
> >>> Cc: Peter Zijlstra <peterz@infradead.org>
> >>> Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
> >>> ---
> >>>    drivers/media/rc/bpf-lirc.c    |  4 ++--
> >>>    include/linux/bpf.h            | 16 +++++++++++++++-
> >>>    include/linux/perf_event.h     |  1 +
> >>>    include/linux/trace_events.h   |  6 +++---
> >>>    include/uapi/linux/bpf.h       |  7 +++++++
> >>>    kernel/bpf/core.c              | 29 ++++++++++++++++++-----------
> >>>    kernel/bpf/syscall.c           |  2 +-
> >>>    kernel/events/core.c           | 21 ++++++++++++++-------
> >>>    kernel/trace/bpf_trace.c       |  8 +++++---
> >>>    tools/include/uapi/linux/bpf.h |  7 +++++++
> >>>    10 files changed, 73 insertions(+), 28 deletions(-)
> >>>
> >>> diff --git a/drivers/media/rc/bpf-lirc.c b/drivers/media/rc/bpf-lirc.c
> >>> index afae0afe3f81..7490494273e4 100644
> >>> --- a/drivers/media/rc/bpf-lirc.c
> >>> +++ b/drivers/media/rc/bpf-lirc.c
> >>> @@ -160,7 +160,7 @@ static int lirc_bpf_attach(struct rc_dev *rcdev, struct bpf_prog *prog)
> >>>                goto unlock;
> >>>        }
> >>>
> >>> -     ret = bpf_prog_array_copy(old_array, NULL, prog, &new_array);
> >>> +     ret = bpf_prog_array_copy(old_array, NULL, prog, 0, &new_array);
> >>>        if (ret < 0)
> >>>                goto unlock;
> >>>
> >> [...]
> >>>    void bpf_trace_run1(struct bpf_prog *prog, u64 arg1);
> >>> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> >>> index 00b1267ab4f0..bc1fd54a8f58 100644
> >>> --- a/include/uapi/linux/bpf.h
> >>> +++ b/include/uapi/linux/bpf.h
> >>> @@ -1448,6 +1448,13 @@ union bpf_attr {
> >>>                                __aligned_u64   iter_info;      /* extra bpf_iter_link_info */
> >>>                                __u32           iter_info_len;  /* iter_info length */
> >>>                        };
> >>> +                     struct {
> >>> +                             /* black box user-provided value passed through
> >>> +                              * to BPF program at the execution time and
> >>> +                              * accessible through bpf_get_user_ctx() BPF helper
> >>> +                              */
> >>> +                             __u64           user_ctx;
> >>> +                     } perf_event;
> >>
> >> Is it possible to fold this field into previous union?
> >>
> >>                   union {
> >>                           __u32           target_btf_id;  /* btf_id of
> >> target to attach to */
> >>                           struct {
> >>                                   __aligned_u64   iter_info;      /*
> >> extra bpf_iter_link_info */
> >>                                   __u32           iter_info_len;  /*
> >> iter_info length */
> >>                           };
> >>                   };
> >>
> >>
> >
> > I didn't want to do it, because different types of BPF links will
> > accept this user_ctx (or now bpf_cookie). And then we'll have to have
> > different locations of that field for different types of links.
> >
> > For example, when/if we add this user_ctx to BPF iterator programs,
> > having __u64 user_ctx in the same anonymous union will make it overlap
> > with iter_info, which is a problem. So I want to have a link
> > type-specific sections in LINK_CREATE command section, to allow the
> > same field name at different locations.
> >
> > I actually think that we should put iter_info/iter_info_len into a
> > named field, like this (also added user_ctx for bpf_iter link as a
> > demonstration):
> >
> > struct {
> >      __aligned_u64 info;
> >      __u32         info_len;
> >      __aligned_u64 user_ctx;  /* see how it's at a different offset
> > than perf_event.user_ctx */
> > } iter;
> > struct {
> >      __u64         user_ctx;
> > } perf_event;
> >
> > (of course keeping already existing fields in anonymous struct for
> > backwards compatibility)
>
> Okay, then since user_ctx may be used by many link types. How
> about just with the field "user_ctx" without struct perf_event.

I'd love to do it because it is indeed generic and common field, like
target_fd. But I'm not sure what you are proposing below. Where
exactly that user_ctx (now called bpf_cookie) goes in your example? I
see few possible options that allow preserving ABI backwards
compatibility. Let's see if you and everyone else likes any of those
better. I'll use the full LINK_CREATE sub-struct definition from
bpf_attr to make it clear. And to demonstrate how this can be extended
to bpf_iter in the future, please note this part as this is an
important aspect.

1. Full backwards compatibility and per-link type sections (my current
approach):

        struct { /* struct used by BPF_LINK_CREATE command */
                __u32           prog_fd;
                union {
                        __u32           target_fd;
                        __u32           target_ifindex;
                };
                __u32           attach_type;
                __u32           flags;
                union {
                        __u32           target_btf_id;
                        struct {
                                __aligned_u64   iter_info;
                                __u32           iter_info_len;
                        };
                        struct {
                                __u64           bpf_cookie;
                        } perf_event;
                        struct {
                                __aligned_u64   info;
                                __u32           info_len;
                                __aligned_u64   bpf_cookie;
                        } iter;
               };
        } link_create;

The good property here is that we can keep easily extending link
type-specific sections with extra fields where needed. For common
stuff like bpf_cookie it's suboptimal because we'll need to duplicate
field definition in each struct inside that union, but I think that's
fine. From end-user point of view, they will know which type of link
they are creating, so the use will be straightforward. This is why I
went with this approach. But let's consider alternatives.

2. Non-backwards compatible layout but extra flag to specify that new
field layout is used.

        struct { /* struct used by BPF_LINK_CREATE command */
                __u32           prog_fd;
                union {
                        __u32           target_fd;
                        __u32           target_ifindex;
                };
                __u32           attach_type;
                __u32           flags; /* this will start supporting
some new flag like BPF_F_LINK_CREATE_NEW */
                __u64           bpf_cookie; /* common field now */
                union { /* this parts is effectively deprecated now */
                        __u32           target_btf_id;
                        struct {
                                __aligned_u64   iter_info;
                                __u32           iter_info_len;
                        };
                        struct { /* this is new layout, but needs
BPF_F_LINK_CREATE_NEW, at least for ext/ and bpf_iter/ programs */
                            __u64       bpf_cookie;
                            union {
                                struct {
                                    __u32     target_btf_id;
                                } ext;
                                struct {
                                    __aligned_u64 info;
                                    __u32         info_len;
                                } iter;
                            }
                        }
                };
        } link_create;

This makes bpf_cookie a common field, but at least for EXT (freplace/)
and ITER (bpf_iter/) links we need to specify extra flag to specify
that we are not using iter_info/iter_info_len/target_btf_id. bpf_iter
then will use iter.info and iter.info_len, and can use plain
bpf_cookie.

IMO, this is way too confusing and a maintainability nightmare.

I'm trying to guess what you are proposing, I can read it two ways,
but let me know if I missed something.

3. Just add bpf_cookie field before link type-specific section.

        struct { /* struct used by BPF_LINK_CREATE command */
                __u32           prog_fd;
                union {
                        __u32           target_fd;
                        __u32           target_ifindex;
                };
                __u32           attach_type;
                __u32           flags;
                __u64           bpf_cookie;  // <<<<<<<<<< HERE
                union {
                        __u32           target_btf_id;
                        struct {
                                __aligned_u64   iter_info;
                                __u32           iter_info_len;
                        };
                };
        } link_create;

This looks really nice and would be great, but that changes offsets
for target_btf_id/iter_info/iter_info_len, so a no go. The only way to
rectify this is what proposal #2 above does with an extra flag.

4. Add bpf_cookie after link-type specific part:

        struct { /* struct used by BPF_LINK_CREATE command */
                __u32           prog_fd;
                union {
                        __u32           target_fd;
                        __u32           target_ifindex;
                };
                __u32           attach_type;
                __u32           flags;
                union {
                        __u32           target_btf_id;
                        struct {
                                __aligned_u64   iter_info;
                                __u32           iter_info_len;
                        };
                        struct {
                };
                __u64           bpf_cookie; // <<<<<<<<<<<<<<<<<< HERE
        } link_create;

This could work. But we are wasting 16 bytes currently used for
target_btf_id/iter_info/iter_info_len. If we later need to do
something link type-specific, we can add it to the existing union if
we need <= 16 bytes, otherwise we'll need to start another union after
bpf_cookie, splitting this into two link type-specific sections.

Overall, this might work, especially assuming we won't need to extend
iter-specific portions. But I really hate that we didn't do named
structs inside that union (i.e., ext.target_btf_id and
iter.info/iter.info_len) and I'd like to rectify that in the follow up
patches with named structs duplicating existing field layout, but with
proper naming. But splitting this LINK_CREATE bpf_attr part into two
unions would make it hard and awkward in the future.

So, thoughts? Did you have something else in mind that I missed?


> Sometime like
>
> __u64   user_ctx;
>
> instead of
>
> struct {
>         __u64   user_ctx;
> } perf_event;
>
> >
> > I decided to not do that in this patch set, though, to not distract
> > from the main goal. But I think we should avoid this shared field
> > "namespace" across different link types going forward.
> >
> >
> >>>                };
> >>>        } link_create;
> >>>
> >> [...]