Re: [PATCH bpf-next v4 1/3] bpf: Parameterize task iterators.

[Kui-Feng Lee <kuifeng@fb.com>]

On Tue, 2022-08-09 at 18:08 -0700, Alexei Starovoitov wrote:
> On Tue, Aug 9, 2022 at 3:35 PM Kui-Feng Lee <kuifeng@fb.com> wrote:
> > 
> > On Tue, 2022-08-09 at 15:12 -0700, Alexei Starovoitov wrote:
> > > On Tue, Aug 9, 2022 at 12:54 PM Kui-Feng Lee <kuifeng@fb.com>
> > > wrote:
> > > > 
> > > > Allow creating an iterator that loops through resources of one
> > > > task/thread.
> > > > 
> > > > People could only create iterators to loop through all
> > > > resources of
> > > > files, vma, and tasks in the system, even though they were
> > > > interested
> > > > in only the resources of a specific task or process.  Passing
> > > > the
> > > > additional parameters, people can now create an iterator to go
> > > > through all resources or only the resources of a task.
> > > > 
> > > > Signed-off-by: Kui-Feng Lee <kuifeng@fb.com>
> > > > ---
> > > >  include/linux/bpf.h            |   8 ++
> > > >  include/uapi/linux/bpf.h       |  36 +++++++++
> > > >  kernel/bpf/task_iter.c         | 134
> > > > +++++++++++++++++++++++++++--
> > > > ----
> > > >  tools/include/uapi/linux/bpf.h |  36 +++++++++
> > > >  4 files changed, 190 insertions(+), 24 deletions(-)
> > > > 
> > > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > > > index 11950029284f..bef81324e5f1 100644
> > > > --- a/include/linux/bpf.h
> > > > +++ b/include/linux/bpf.h
> > > > @@ -1718,6 +1718,14 @@ int bpf_obj_get_user(const char __user
> > > > *pathname, int flags);
> > > > 
> > > >  struct bpf_iter_aux_info {
> > > >         struct bpf_map *map;
> > > > +       struct {
> > > > +               enum bpf_iter_task_type type;
> > > > +               union {
> > > > +                       u32 tid;
> > > > +                       u32 tgid;
> > > > +                       u32 pid_fd;
> > > > +               };
> > > > +       } task;
> > > >  };
> > > > 
> > > >  typedef int (*bpf_iter_attach_target_t)(struct bpf_prog *prog,
> > > > diff --git a/include/uapi/linux/bpf.h
> > > > b/include/uapi/linux/bpf.h
> > > > index ffcbf79a556b..3d0b9e34089f 100644
> > > > --- a/include/uapi/linux/bpf.h
> > > > +++ b/include/uapi/linux/bpf.h
> > > > @@ -87,10 +87,46 @@ struct bpf_cgroup_storage_key {
> > > >         __u32   attach_type;            /* program attach type
> > > > (enum bpf_attach_type) */
> > > >  };
> > > > 
> > > > +/*
> > > > + * The task type of iterators.
> > > > + *
> > > > + * For BPF task iterators, they can be parameterized with
> > > > various
> > > > + * parameters to visit only some of tasks.
> > > > + *
> > > > + * BPF_TASK_ITER_ALL (default)
> > > > + *     Iterate over resources of every task.
> > > > + *
> > > > + * BPF_TASK_ITER_TID
> > > > + *     Iterate over resources of a task/tid.
> > > > + *
> > > > + * BPF_TASK_ITER_TGID
> > > > + *     Iterate over reosurces of evevry task of a process /
> > > > task
> > > > group.
> > > > + *
> > > > + * BPF_TASK_ITER_PIDFD
> > > > + *     Iterate over resources of every task of a process /task
> > > > group specified by a pidfd.
> > > > + */
> > > > +enum bpf_iter_task_type {
> > > > +       BPF_TASK_ITER_ALL = 0,
> > > > +       BPF_TASK_ITER_TID,
> > > > +       BPF_TASK_ITER_TGID,
> > > > +       BPF_TASK_ITER_PIDFD,
> > > > +};
> > > > +
> > > >  union bpf_iter_link_info {
> > > >         struct {
> > > >                 __u32   map_fd;
> > > >         } map;
> > > > +       /*
> > > > +        * Parameters of task iterators.
> > > > +        */
> > > > +       struct {
> > > > +               enum bpf_iter_task_type type;
> > > > +               union {
> > > > +                       __u32 tid;
> > > > +                       __u32 tgid;
> > > > +                       __u32 pid_fd;
> > > > +               };
> > > 
> > > Sorry I'm late to this discussion, but
> > > with enum and with union we kinda tell
> > > the kernel the same information twice.
> > > Here is how the selftest looks:
> > > +       linfo.task.tid = getpid();
> > > +       linfo.task.type = BPF_TASK_ITER_TID;
> > > 
> > > first line -> use tid.
> > > second line -> yeah. I really meant the tid.
> > > 
> > > Instead of union and type can we do:
> > > > +                       __u32 tid;
> > > > +                       __u32 tgid;
> > > > +                       __u32 pid_fd;
> > > 
> > > as 3 separate fields?
> > > The kernel would have to check that only one
> > > of them is set.
> > > 
> > > I could have missed an earlier discussion on this subj.
> > 
> > We may have other parameter types later, for example, cgroups.
> > Unfortunately, we don't have tagged enum or tagged union, like what
> > Rust or Haskell has, in C.  A separated 'type' field would be
> > easier
> > and clear to distinguish them.  With 3 separated fields, people may
> > wonder if they can be set them all, or just one of them, in my
> > opinion.
> > With an union, people should know only one of them should be set.
> 
> What stops us adding new fields to the end in such a case?
> Some combinations will not be meaningful and the kernel
> would have to check and error regardless.
> Imagine extending union:
> struct {
>   enum bpf_iter_task_type type;
>   union {
>      struct {
>         __u32 tid;
>         __u64 something_else;
>      };
>      __u32 tgid;
>      __u32 pid_fd;
>   };
> };
> 
> and now we're suddenly hitting the same issue we discussed
> with struct bpf_link_info in the other thread due to alignment
> increasing from 4 to 8 bytes.
> We might even need bpf_iter_link_info _v2.

It is a good point.  In that case, we probably need task_v2 instead of
bpf_iter_link_info_v2.  The other solution is to make whole 'task' as
an union instead of a struct.

union bpf_iter_link_info {
    ......
    union {
        enum bpf_iter_task_type type;
        struct {
            enum bpf_iter_task_type tid__type;
            __u32 tid;
        };
        struct {
            enum bpf_iter_task_type tgid__type;
            __u32 tgid;
        };
        ......
    } task;
}

Even adding something new, it doesn't affect the offsets of old fields.

> 
> If 'something_else' is u32 the kernel still needs to check
> that it's zero in the tgid and pid_fd cases.
> If we're extending fields we can add a comment:
> struct {
>   __u32 tid;
>   __u32 tgid;
>   __u32 pid_fd;
>   __u32 something_else; /* useful in combination with tid */
> };
> and it's obvious what is used with what.
> 
> It still feels that 3 different fields are easier to use.

Agree!  Having 3 separated fields is easier to use for assigning only
one value.