[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CAEf4BzYeMBYDx6TXAqdkpWpW43yKWNbGaA+67LV1zPao-u779A@mail.gmail.com>
Date: Thu, 24 Oct 2019 20:19:56 -0700
From: Andrii Nakryiko <andrii.nakryiko@...il.com>
To: Toke Høiland-Jørgensen <toke@...hat.com>
Cc: Daniel Borkmann <daniel@...earbox.net>,
Alexei Starovoitov <ast@...nel.org>,
Martin KaFai Lau <kafai@...com>,
Song Liu <songliubraving@...com>, Yonghong Song <yhs@...com>,
Jesper Dangaard Brouer <brouer@...hat.com>,
David Miller <davem@...emloft.net>,
Networking <netdev@...r.kernel.org>, bpf <bpf@...r.kernel.org>
Subject: Re: [PATCH bpf-next v2 3/4] libbpf: Support configurable pinning of
maps from BTF annotations
On Thu, Oct 24, 2019 at 6:11 AM Toke Høiland-Jørgensen <toke@...hat.com> wrote:
>
> From: Toke Høiland-Jørgensen <toke@...hat.com>
>
> This adds support to libbpf for setting map pinning information as part of
> the BTF map declaration. We introduce a version new
> bpf_object__map_pin_opts() function to pin maps based on this setting, as
> well as a getter and setter function for the pin information that callers
> can use after map load.
>
> The pinning type currently only supports a single PIN_BY_NAME mode, where
> each map will be pinned by its name in a path that can be overridden, but
> defaults to /sys/fs/bpf.
>
> The pinning options supports a 'pin_all' setting, which corresponds to the
> old bpf_object__map_pin() function with an explicit path. In addition, the
> function now defaults to just skipping over maps that are already
> pinned (since the previous commit started recording this in struct
> bpf_map). This behaviour can be turned off with the 'no_skip_pinned' option.
>
> Signed-off-by: Toke Høiland-Jørgensen <toke@...hat.com>
> ---
I think you are overcomplicating this... Here's how I think we can
satisfy both simplicity goals, as well as good usability:
1. add `const char *pin_root_path` to bpf_object_open_opts. This
pinning path override doesn't need to leave in some separate set of
options, it's BPF object's parameter, so let's put it into open
settings.
2. If BTF-defined map definition has pinning set to PIN_BY_NAME, that
means bpf_object__load should do auto-pinning. If not, no
auto-pinning, only if manually requested by explicit bpf_map__pin.
Further, if someone wants to auto-pin map to a custom location, do
bpf_map__set_pin_path() before bpf_object__load(), and load should
auto-pin it as well.
3. bpf_map__get_pinning/bpf_map__set_pinning are unnecessary, at least
for now. Let's not add unnecessary APIs.
4. pin_all/skip_pinned seems unnecessary. What scenarios are you
solving with them? Given #1 and #4, just drop
bpf_object__pin_maps_opts().
The way I see it, libbpf should behave sanely for declarative use
case, but allow custom tuning programmatically. If map is set to
PIN_BY_NAME in map definition - we derive pin_path (potentially taking
into account custom pin root path from open opts) and auto-pin on load
(unless application set pin_path manually). In a weird case, where map
is declaratively defined as auto-pinnable, but application for
whatever reason decides not to do it - it can unset pin_path with
bpf_map__set_pin_path(NULL).
Full control, but simple and intuitive default behavior? Does it make sense?
> tools/lib/bpf/bpf_helpers.h | 6 ++
> tools/lib/bpf/libbpf.c | 134 ++++++++++++++++++++++++++++++++++---------
> tools/lib/bpf/libbpf.h | 26 ++++++++
> tools/lib/bpf/libbpf.map | 3 +
> 4 files changed, 142 insertions(+), 27 deletions(-)
>
Powered by blists - more mailing lists