On Mon, Feb 8, 2021 at 7:55 PM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote:
>
> Gilad Reti <gilad.reti@xxxxxxxxx> writes:
>
> > On Mon, Feb 8, 2021 at 5:09 PM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote:
> >>
> >> Gilad Reti <gilad.reti@xxxxxxxxx> writes:
> >>
> >> > On Mon, Feb 8, 2021 at 4:28 PM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote:
> >> >>
> >> >> Gilad Reti <gilad.reti@xxxxxxxxx> writes:
> >> >>
> >> >> > On Mon, Feb 8, 2021 at 1:42 PM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote:
> >> >> >>
> >> >> >> Gilad Reti <gilad.reti@xxxxxxxxx> writes:
> >> >> >>
> >> >> >> > Also, is there a way to set the pin path to all maps/programs at once?
> >> >> >> > For example, bpf_object__pin_maps pins all maps at a specific path,
> >> >> >> > but as far as I was able to find there is no similar function to set
> >> >> >> > the pin path for all maps only (without pinning) so that at loading
> >> >> >> > time libbpf will try to reuse all maps. The only way to achieve a
> >> >> >> > complete reuse of all maps that I could find is to "reverse engineer"
> >> >> >> > libbpf's pin path generation algorithm (i.e. <path>/<map_name>) and
> >> >> >> > set the pin path on each map before load.
> >> >> >>
> >> >> >> You can set the 'pinning' attribute in the map definition - add
> >> >> >> '__uint(pinning, LIBBPF_PIN_BY_NAME);' to the map struct. By default
> >> >> >> this will pin beneath /sys/fs/bpf, but you can customise that by setting
> >> >> >> the pin_root_path attribute in bpf_object_open_opts.
> >> >> >
> >> >> > Yes, I am familiar with that feature, but it has some downsides:
> >> >> > 1. I need to set it manually on every map (and in cases that I have
> >> >> > only the compiled object file that would be hard).
> >> >> > 2. It only works for bpf maps and not bpf programs.
> >> >> > 3. It only works for bpf maps that are defined explicitly in the bpf
> >> >> > code and not for implicit (inner) bpf maps (bss, rodata, etc).
> >> >>
> >> >> Ah, right. Well, other than that I don't think there's a way to set pin
> >> >> paths in bulk, other than by manually iterating and setting them one at
> >> >> a time. But, erm, can't you just do that? :)
> >> >>
> >> >
> >> > Sure, I can, but I think we should avoid that. As I said this forces
> >> > the user to know libbpf's pin path naming algorithm, which is not part
> >> > of the libbpf api afaik.
> >>
> >> Why? If you set the pin path from your application, libbpf will also try
> >> to reuse the map from that path. So you don't need to know libbpf's
> >> algorithm if you just override it with your own paths?
> >>
> >
> > If I do bpf_object__pin_maps then libbpf decides where it wants to pin
> > them. I can set each path by my own, but then why do we need this
> > function?
>
> Erm, what do you mean, "libbpf decides". bpf_object__pin_maps(obj, path)
> does exactly what you're asking for: If you supply the path, all maps
> are going to be pinned by name underneath that directory...
They are pinned under this directory, but with which filename? Today
libbpf builds the filename by taking the map name and escaping it, but
what will happen if this will have to change?
For example, libbpf pins programs by thier section name, but as I
mentioned in the first message in this thread this causes failures
when multiple programs reside in one section, so we may need to change
the naming and this can break user code.
>
> > All It does is pinning all the maps at paths that are not part of the
> > api. In this libbpf version it is here, in the next it is there, and
> > user code will need to change accordingly.
>
> Why would you assume the paths would change? That would be an API break?
The point that I am trying to make is that this naming convention is
libbpf internal, so either it should be defined as a part of the api
and each user will need to implement the same naming logic on his side
(escaping etc), relying on libbpf implementation to remain consistent
(this may be the case right now) but for in cases like the program
pinning for example api will have to break, or just expose the pinning
path as another api and then libbpf will need to make no guarantees
about his naming convention.
>
> > For example, libbpf today uses <path>/<map_name> as the pin path, but
> > it is also doing sanitize_pin_path on each path. This means that after
> > if use bpf_object__pin_maps I also need to know how libbpf sanitizes
> > its paths and mimic that behavior on my side.
>
> The paths are sanitised so the kernel will accept them. If you're using
> invalid paths your pinning is not going to work at all. If you just want
> the paths that the maps are pinned under, use bpf_map__get_pin_path().
>
I am not saying that sanitization is redundant, but rather that it
needs to be properly defined (i.e. all dots will always be replaced
with underscores), so either expose it in the api or document it so
that users don't have to look after the specific implementation.
Of course I can retrieve that paths after pinning, but this is sounds
like sort of a workaround to me, though I am okay with staying this
way. The problem that I have with this approach is that if other
processes want to reuse the same map they will have to share the
pinning path somehow instead of each one retreiving it from the map
directly.
> >> > I think that if we have a method to pin all maps at a specific path
> >> > there should also be a method for reusing them all from this path,
> >> > either by exposing the function that builds the pin path, or a
> >> > function that sets all the paths from a root path.
> >>
> >> What you're asking for is basically a function
> >> bpf_object__set_all_pin_paths(obj, path)
> >>
> >> instead of having to do
> >>
> >> bpf_object__for_each_map(map, obj) {
> >> sprintf(path, "path/%s", bpf_map__name(map));
> >> bpf_map__set_pin_path(map, path);
> >> }
> >>
> >> or? Is that really needed?
> >>
> >
> > Yes, that is what I am asking for. Either that or a
> > bpf_map__build_pin_path(path, map) That will return a pin path that is
> > compatible with libbpf's one, and then I can iterate over all maps.
>
> See above; this is what bpf_object__pin_maps() does today?
I didn't get this last comment. What I meant is that I want something
like the bpf_object__pin_maps but that doesn't pin the maps, just
exposing its naming part.
>
> -Toke
>
