On Tue, Apr 20, 2021 at 8:58 AM Daniel Borkmann [off-list ref] wrote:

On 4/20/21 3:17 AM, Alexei Starovoitov wrote:

quoted

On Thu, Apr 15, 2021 at 10:47 AM Pedro Tammela [off-list ref] wrote:

quoted

Andrii suggested to remove this abstraction layer and have the percpu
handling more explicit[1].

This patch also updates the tests that relied on the macros.

[1] https://lore.kernel.org/bpf/CAEf4BzYmj_ZPDq8Zi4dbntboJKRPU2TVopysBNrdd9foHTfLZw@mail.gmail.com/ (local)

Suggested-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Pedro Tammela <redacted>
---
  tools/testing/selftests/bpf/bpf_util.h        |  7 --
  .../bpf/map_tests/htab_map_batch_ops.c        | 87 +++++++++----------
  .../selftests/bpf/prog_tests/map_init.c       |  9 +-
  tools/testing/selftests/bpf/test_maps.c       | 84 +++++++++++-------
  4 files changed, 96 insertions(+), 91 deletions(-)

diff --git a/tools/testing/selftests/bpf/bpf_util.h b/tools/testing/selftests/bpf/bpf_util.h
index a3352a64c067..105db3120ab4 100644
--- a/tools/testing/selftests/bpf/bpf_util.h
+++ b/tools/testing/selftests/bpf/bpf_util.h

@@ -20,13 +20,6 @@ static inline unsigned int bpf_num_possible_cpus(void)
         return possible_cpus;
  }

-#define __bpf_percpu_val_align __attribute__((__aligned__(8)))
-
-#define BPF_DECLARE_PERCPU(type, name)                         \
-       struct { type v; /* padding */ } __bpf_percpu_val_align \
-               name[bpf_num_possible_cpus()]
-#define bpf_percpu(name, cpu) name[(cpu)].v
-

Hmm. I wonder what Daniel has to say about it, since he
introduced it in commit f3515b5d0b71 ("bpf: provide a generic macro
for percpu values for selftests")
to address a class of bugs.

I would probably even move those into libbpf instead. ;-) The problem is that this can
be missed easily and innocent changes would lead to corruption of the applications
memory if there's a map lookup. Having this at least in selftest code or even in libbpf
would document code-wise that care needs to be taken on per cpu maps. Even if we'd put
a note under Documentation/bpf/ or such, this might get missed easily and finding such
bugs is like looking for a needle in a haystack.. so I don't think this should be removed.

See [0] for previous discussion. I don't mind leaving bpf_percpu() in
selftests. I'm not sure I ever suggested removing it from selftests,
but I don't think it's a good idea to add it to libbpf. I think it's
better to have an extra paragraph in bpf_lookup_map_elem() in
uapi/linux/bpf.h mentioning how per-CPU values should be read/updated.
I think we should just recommend to use u64 for primitive values (or
otherwise users can embed their int in custom aligned(8) struct, if
they insist on <u64) and __attribute__((aligned(8))) for structs.

  [0] https://lore.kernel.org/bpf/CAEf4BzaLKm_fy4oO4Rdp76q2KoC6yC1WcJLuehoZUu9JobG-Cw@mail.gmail.com/ (local)

Thanks,
Daniel