Re: [PATCH bpf-next 1/2] bpf, libbpf: add a new API bpf_object__reuse_maps()

[Anton Protopopov <a.s.protopopov@gmail.com>]

вт, 9 июл. 2019 г. в 13:40, Andrii Nakryiko <andrii.nakryiko@gmail.com>:
>
> On Mon, Jul 8, 2019 at 1:37 PM Anton Protopopov
> <a.s.protopopov@gmail.com> wrote:
> >
> > пн, 8 июл. 2019 г. в 13:54, Andrii Nakryiko <andrii.nakryiko@gmail.com>:
> > >
> > > On Fri, Jul 5, 2019 at 2:53 PM Daniel Borkmann <daniel@iogearbox.net> wrote:
> > > >
> > > > On 07/05/2019 10:44 PM, Anton Protopopov wrote:
> > > > > Add a new API bpf_object__reuse_maps() which can be used to replace all maps in
> > > > > an object by maps pinned to a directory provided in the path argument.  Namely,
> > > > > each map M in the object will be replaced by a map pinned to path/M.name.
> > > > >
> > > > > Signed-off-by: Anton Protopopov <a.s.protopopov@gmail.com>
> > > > > ---
> > > > >  tools/lib/bpf/libbpf.c   | 34 ++++++++++++++++++++++++++++++++++
> > > > >  tools/lib/bpf/libbpf.h   |  2 ++
> > > > >  tools/lib/bpf/libbpf.map |  1 +
> > > > >  3 files changed, 37 insertions(+)
> > > > >
> > > > > diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
> > > > > index 4907997289e9..84c9e8f7bfd3 100644
> > > > > --- a/tools/lib/bpf/libbpf.c
> > > > > +++ b/tools/lib/bpf/libbpf.c
> > > > > @@ -3144,6 +3144,40 @@ int bpf_object__unpin_maps(struct bpf_object *obj, const char *path)
> > > > >       return 0;
> > > > >  }
> > > > >
> > > > > +int bpf_object__reuse_maps(struct bpf_object *obj, const char *path)
> > >
> > > As is, bpf_object__reuse_maps() can be easily implemented by user
> > > applications, as it's only using public libbpf APIs, so I'm not 100%
> > > sure we need to add method like that to libbpf.
> >
> > The bpf_object__reuse_maps() can definitely be implemented by user
> > applications, however, to use it a user also needs to re-implement the
> > bpf_prog_load_xattr funciton, so it seemed to me that adding this
> > functionality to the library is a better way.
>
> I'm still not convinced. Looking at bpf_prog_load_xattr, I think some
> of what it's doing should be part of bpf_object__object_xattr anyway
> (all the expected type setting for programs).
>
> Besides that, there isn't much more than just bpf_object__open and
> bpf_object__load, to be honest. By doing open and load explicitly,
> user gets an opportunity to do whatever adjustment they need: reuse
> maps, adjust map sizes, etc. So I think we should improve
> bpf_object__open to "guess" program attach types and add map
> definition flags to allow reuse declaratively.
>
>
> >
> > >
> > > > > +{
> > > > > +     struct bpf_map *map;
> > > > > +
> > > > > +     if (!obj)
> > > > > +             return -ENOENT;
> > > > > +
> > > > > +     if (!path)
> > > > > +             return -EINVAL;
> > > > > +
> > > > > +     bpf_object__for_each_map(map, obj) {
> > > > > +             int len, err;
> > > > > +             int pinned_map_fd;
> > > > > +             char buf[PATH_MAX];
> > > >
> > > > We'd need to skip the case of bpf_map__is_internal(map) since they are always
> > > > recreated for the given object.
> > > >
> > > > > +             len = snprintf(buf, PATH_MAX, "%s/%s", path, bpf_map__name(map));
> > > > > +             if (len < 0) {
> > > > > +                     return -EINVAL;
> > > > > +             } else if (len >= PATH_MAX) {
> > > > > +                     return -ENAMETOOLONG;
> > > > > +             }
> > > > > +
> > > > > +             pinned_map_fd = bpf_obj_get(buf);
> > > > > +             if (pinned_map_fd < 0)
> > > > > +                     return pinned_map_fd;
> > > >
> > > > Should we rather have a new map definition attribute that tells to reuse
> > > > the map if it's pinned in bpf fs, and if not, we create it and later on
> > > > pin it? This is what iproute2 is doing and which we're making use of heavily.
> > >
> > > I'd like something like that as well. This would play nicely with
> > > recently added BTF-defined maps as well.
> > >
> > > I think it should be not just pin/don't pin flag, but rather pinning
> > > strategy, to accommodate various typical strategies of handling maps
> > > that are already pinned. So something like this:
> > >
> > > 1. BPF_PIN_NOTHING - default, don't pin;
> > > 2. BPF_PIN_EXCLUSIVE - pin, but if map is already pinned - fail;
> > > 3. BPF_PIN_SET - pin; if existing map exists, reset its state to be
> > > exact state of object's map;
> > > 4. BPF_PIN_MERGE - pin, if map exists, fill in NULL entries only (this
> > > is how Cilium is pinning PROG_ARRAY maps, if I understand correctly);
> > > 5. BPF_PIN_MERGE_OVERWRITE - pin, if map exists, overwrite non-NULL values.
> > >
> > > This list is only for illustrative purposes, ideally people that have
> > > a lot of experience using pinning for real-world use cases would chime
> > > in on what strategies are useful and make sense.
> >
> > My case was simply to reuse existing maps when reloading a program.
> > Does it make sense for you to add only the simplest cases of listed above?
>
> Of course, it's enum, so we can start with few clearly useful ones and
> then expand more if we ever have a need. But I think we still need a
> bit wider discussion and let people who use pinning to chime in.
>
> >
> > Also, libbpf doesn't use standard naming conventions for pinning maps.
>
> We talked about this in another thread related to BTF-defined maps. I
> think the way to go with this is to actually define a default pinning
> root path, but allow to override it on bpf_object__open, if user needs
> a different one.
>
> > Does it make sense to provide a list of already open maps to the
> > bpf_prog_load_xattr function as an attribute? In this case a user
> > can execute his own policy on pinning, but still will have an option
> > to reuse, reset, and merge maps.
>
> As explained above, I don't think there isn't much added value in
> bpf_prog_load, so I'd advise to just switch to explicit
> bpf_object__open + bpf_object__load and get maximum control and
> flexibility.

Thanks for your comments. I can see now that using
bpf_object__open/bpf_object__load makes better sense.

>
> >
> > >
> > > > In bpf_object__reuse_maps() bailing out if bpf_obj_get() fails is perhaps
> > > > too limiting for a generic API as new version of an object file may contain
> > > > new maps which are not yet present in bpf fs at that point.
> > > >
> > > > > +             err = bpf_map__reuse_fd(map, pinned_map_fd);
> > > > > +             if (err)
> > > > > +                     return err;
> > > > > +     }
> > > > > +
> > > > > +     return 0;
> > > > > +}
> > > > > +
> > > > >  int bpf_object__pin_programs(struct bpf_object *obj, const char *path)
> > > > >  {
> > > > >       struct bpf_program *prog;
> > > > > diff --git a/tools/lib/bpf/libbpf.h b/tools/lib/bpf/libbpf.h
> > > > > index d639f47e3110..7fe465a1be76 100644
> > > > > --- a/tools/lib/bpf/libbpf.h
> > > > > +++ b/tools/lib/bpf/libbpf.h
> > > > > @@ -82,6 +82,8 @@ int bpf_object__variable_offset(const struct bpf_object *obj, const char *name,
> > > > >  LIBBPF_API int bpf_object__pin_maps(struct bpf_object *obj, const char *path);
> > > > >  LIBBPF_API int bpf_object__unpin_maps(struct bpf_object *obj,
> > > > >                                     const char *path);
> > > > > +LIBBPF_API int bpf_object__reuse_maps(struct bpf_object *obj,
> > > > > +                                   const char *path);
> > > > >  LIBBPF_API int bpf_object__pin_programs(struct bpf_object *obj,
> > > > >                                       const char *path);
> > > > >  LIBBPF_API int bpf_object__unpin_programs(struct bpf_object *obj,
> > > > > diff --git a/tools/lib/bpf/libbpf.map b/tools/lib/bpf/libbpf.map
> > > > > index 2c6d835620d2..66a30be6696c 100644
> > > > > --- a/tools/lib/bpf/libbpf.map
> > > > > +++ b/tools/lib/bpf/libbpf.map
> > > > > @@ -172,5 +172,6 @@ LIBBPF_0.0.4 {
> > > > >               btf_dump__new;
> > > > >               btf__parse_elf;
> > > > >               bpf_object__load_xattr;
> > > > > +             bpf_object__reuse_maps;
> > > > >               libbpf_num_possible_cpus;
> > > > >  } LIBBPF_0.0.3;
> > > > >
> > > >