public inbox for bpf@vger.kernel.org
 help / color / mirror / Atom feed
* [RFC v2 PATCH bpf-next 0/4] bpf: add percpu stats for bpf_map
@ 2023-06-22  9:53 Anton Protopopov
  2023-06-22  9:53 ` [RFC v2 PATCH bpf-next 1/4] bpf: add percpu stats for bpf_map elements insertions/deletions Anton Protopopov
                   ` (4 more replies)
  0 siblings, 5 replies; 15+ messages in thread
From: Anton Protopopov @ 2023-06-22  9:53 UTC (permalink / raw)
  To: Alexei Starovoitov, Daniel Borkmann, John Fastabend,
	Andrii Nakryiko, Martin KaFai Lau, Song Liu, Yonghong Song,
	KP Singh, Stanislav Fomichev, Hao Luo, Jiri Olsa, bpf
  Cc: Anton Protopopov

This series adds a mechanism for maps to populate per-cpu counters of elements
on insertions/deletions. The sum of these counters can be accessed by a new
kfunc from a map iterator program.

The following patches are present in the series:

  * Patch 1 adds a generic per-cpu counter to struct bpf_map
  * Patch 2 utilizes this mechanism for hash-based maps
  * Patch 3 extends the preloaded map iterator to dump the sum
  * Patch 4 adds a self-test for the change

The reason for adding this functionality in our case (Cilium) is to get
signals about how full some heavy-used maps are and what the actual dynamic
profile of map capacity is. In the case of LRU maps this is impossible to get
this information anyhow else. See also [1].

This is a v2 for the https://lore.kernel.org/bpf/20230531110511.64612-1-aspsk@isovalent.com/T/#t
It was rewritten according to previous comments.  I've turned this series into
an RFC for two reasons:

1) This patch only works on systems where this_cpu_{inc,dec} is atomic for s64.
For systems which might write s64 non-atomically this would be required to use
some locking mechanism to prevent readers from reading trash via the
bpf_map_sum_elements_counter() kfunc (see patch 1)

2) In comparison with the v1, we're adding extra instructions per map operation
(for preallocated maps, as well as for non-preallocated maps). The only
functionality we're interested at the moment is the number of elements present
in a map, not a per-cpu statistics. This could be better achieved by using
the v1 version, which only adds computations for preallocated maps.

So, the question is: won't it be fine to do the changes in the following way:

  * extend the preallocated hash maps to populate percpu batch counters as in v1
  * add a kfunc as in v2 to get the current sum

This works as

  * nobody at the moment actually requires the per-cpu statistcs
  * this implementation can be transparently turned into per-cpu statistics, if
    such a need occurs on practice (the only thing to change would be to
    re-implement the kfunc and, maybe, add more kfuncs to get per-cpu stats)
  * the "v1 way" is the least intrusive: it only affects preallocated maps, as
    other maps already provide the required functionality

  [1] https://lpc.events/event/16/contributions/1368/

v1 -> v2:
- make the counters generic part of struct bpf_map
- don't use map_info and /proc/self/fdinfo in favor of a kfunc

Anton Protopopov (4):
  bpf: add percpu stats for bpf_map elements insertions/deletions
  bpf: populate the per-cpu insertions/deletions counters for hashmaps
  bpf: make preloaded map iterators to display map elements count
  selftests/bpf: test map percpu stats

 include/linux/bpf.h                           |  30 +
 kernel/bpf/hashtab.c                          | 102 ++--
 kernel/bpf/map_iter.c                         |  48 +-
 kernel/bpf/preload/iterators/iterators.bpf.c  |   9 +-
 .../iterators/iterators.lskel-little-endian.h | 513 +++++++++---------
 .../bpf/map_tests/map_percpu_stats.c          | 336 ++++++++++++
 .../selftests/bpf/progs/map_percpu_stats.c    |  24 +
 7 files changed, 766 insertions(+), 296 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/map_tests/map_percpu_stats.c
 create mode 100644 tools/testing/selftests/bpf/progs/map_percpu_stats.c

-- 
2.34.1


^ permalink raw reply	[flat|nested] 15+ messages in thread
end of thread, other threads:[~2023-06-26 14:36 UTC | newest]

Thread overview: 15+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2023-06-22  9:53 [RFC v2 PATCH bpf-next 0/4] bpf: add percpu stats for bpf_map Anton Protopopov
2023-06-22  9:53 ` [RFC v2 PATCH bpf-next 1/4] bpf: add percpu stats for bpf_map elements insertions/deletions Anton Protopopov
2023-06-22 20:11   ` Alexei Starovoitov
2023-06-23 12:47     ` Anton Protopopov
2023-06-23 10:51   ` Daniel Borkmann
2023-06-23 12:35     ` Anton Protopopov
2023-06-22  9:53 ` [RFC v2 PATCH bpf-next 2/4] bpf: populate the per-cpu insertions/deletions counters for hashmaps Anton Protopopov
2023-06-22 20:18   ` Alexei Starovoitov
2023-06-22  9:53 ` [RFC v2 PATCH bpf-next 3/4] bpf: make preloaded map iterators to display map elements count Anton Protopopov
2023-06-22  9:58 ` [RFC v2 PATCH bpf-next 4/4] selftests/bpf: test map percpu stats Anton Protopopov
2023-06-22 20:20   ` Alexei Starovoitov
2023-06-26 14:37     ` Anton Protopopov
2023-06-23  9:53 ` [RFC v2 PATCH bpf-next 0/4] bpf: add percpu stats for bpf_map Daniel Borkmann
2023-06-24  0:17   ` Alexei Starovoitov
2023-06-26  8:50     ` Daniel Borkmann

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox