Message-ID: <CAKY_9u1iyyuJaYz3csoHnW_apvB=hWBosvQm+jxm6Gc4yFM9kA@mail.gmail.com>
Date:   Wed, 7 Apr 2021 17:14:56 -0300
From:   Pedro Tammela <pctammela@...il.com>
To:     Andrii Nakryiko <andrii.nakryiko@...il.com>
Cc:     Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Andrii Nakryiko <andrii@...nel.org>,
        Martin KaFai Lau <kafai@...com>,
        Song Liu <songliubraving@...com>, Yonghong Song <yhs@...com>,
        John Fastabend <john.fastabend@...il.com>,
        KP Singh <kpsingh@...nel.org>, Shuah Khan <shuah@...nel.org>,
        Pedro Tammela <pctammela@...atatu.com>,
        Matthieu Baerts <matthieu.baerts@...sares.net>,
        David Verbeiren <david.verbeiren@...sares.net>,
        "open list:BPF (Safe dynamic programs and tools)" 
        <netdev@...r.kernel.org>,
        "open list:BPF (Safe dynamic programs and tools)" 
        <bpf@...r.kernel.org>, open list <linux-kernel@...r.kernel.org>,
        "open list:KERNEL SELFTEST FRAMEWORK" 
        <linux-kselftest@...r.kernel.org>
Subject: Re: [PATCH bpf-next v2 2/3] libbpf: selftests: refactor
 'BPF_PERCPU_TYPE()' and 'bpf_percpu()' macros

Em qua., 7 de abr. de 2021 às 16:51, Andrii Nakryiko
<andrii.nakryiko@...il.com> escreveu:
>
> On Wed, Apr 7, 2021 at 12:30 PM Pedro Tammela <pctammela@...il.com> wrote:
> >
> > Em qua., 7 de abr. de 2021 às 15:31, Andrii Nakryiko
> > <andrii.nakryiko@...il.com> escreveu:
> > >
> > > On Tue, Apr 6, 2021 at 11:55 AM Pedro Tammela <pctammela@...il.com> wrote:
> > > >
> > > > This macro was refactored out of the bpf selftests.
> > > >
> > > > Since percpu values are rounded up to '8' in the kernel, a careless
> > > > user in userspace might encounter unexpected values when parsing the
> > > > output of the batched operations.
> > >
> > > I wonder if a user has to be more careful, though? This
> > > BPF_PERCPU_TYPE, __bpf_percpu_align and bpf_percpu macros seem to
> > > create just another opaque layer. It actually seems detrimental to me.
> > >
> > > I'd rather emphasize in the documentation (e.g., in
> > > bpf_map_lookup_elem) that all per-cpu maps are aligning values at 8
> > > bytes, so user has to make sure that array of values provided to
> > > bpf_map_lookup_elem() has each element size rounded up to 8.
> >
> > From my own experience, the documentation has been a very unreliable
> > source, to the point that I usually jump to the code first rather than
> > to the documentation nowadays[1].
>
> I totally agree, which is why I think improving docs is necessary.
> Unfortunately docs are usually lagging behind, because generally
> people hate writing documentation and it's just a fact of life.
>
> > Tests, samples and projects have always been my source of truth and we
> > are already lacking a bit on those as well. For instance, the samples
> > directory contains programs that are very outdated (I didn't check if
> > they are still functional).
>
> Yeah, samples/bpf is bitrotting. selftests/bpf, though, are maintained
> and run regularly and vigorously, so making sure they set a good and
> realistic example is a good.
>
>
> > I think macros like these will be present in most of the project
> > dealing with batched operations and as a daily user of libbpf I don't
> > see how this could not be offered by libbpf as a standardized way to
> > declare percpu types.
>
> If I were using per-CPU maps a lot, I'd make sure I use u64 and
> aligned(8) types and bypass all the macro ugliness, because there is
> no need in it and it just hurts readability. So I don't want libbpf to
> incentivize bad choices here by providing seemingly convenient macros.
> Users have to be aware that values are 8-byte aligned/extended. That's
> not a big secret and not a very obscure thing to learn anyways.
>
> >
> > [1] So batched operations were introduced a little bit over a 1 year
> > ago and yet the only reference I had for it was the selftests. The
> > documentation is on my TODO list, but that's just because I have to
> > deal with it daily.
> >
>
> Yeah, please do contribute them!
>
> > >
> > > In practice, I'd recommend users to always use __u64/__s64 when having
> > > primitive integers in a map (they are not saving anything by using
> > > int, it just creates an illusion of savings). Well, maybe on 32-bit
> > > arches they would save a bit of CPU, but not on typical 64-bit
> > > architectures. As for using structs as values, always mark them as
> > > __attribute__((aligned(8))).
> > >
> > > Basically, instead of obscuring the real use some more, let's clarify
> > > and maybe even provide some examples in documentation?
> >
> > Why not do both?
> >
> > Provide a standardized way to declare a percpu value with examples and
> > a good documentation with examples.
> > Let the user decide what is best for his use case.
>
> What is a standardized way? A custom macro with struct { T v; }
> inside? That's just one way of doing this, and it requires another
> macro to just access the value (because no one wants to write
> my_values[cpu].v, right?). I'd say the standardized way of reading
> values should look like `my_values[cpu]`, that's it. For that you use
> 64-bit integers or 8-byte aligned structs. And don't mess with macros
> for that at all.
>
> So if a user insists on using int/short/char as value, they can do
> their own struct { char v} __aligned(8) trick. But I'd advise such
> users to reconsider and use u64. If they are using structs for values,
> always mark __aligned(8) and forget about this in the rest of your
> code.
>
> As for allocating memory for array of per-cpu values, there is also no
> single standardized way we can come up with. It could be malloc() on
> the heap. Or alloca() on the stack. Or it could be pre-allocated one
> for up to maximum supported CPUs. Or... whatever makes sense.
>
> So I think the best way to handle all that is to clearly explain how
> reading per-CPU values from per-CPU maps works in BPF and what are the
> memory layout expectations.

I understand your points much better now. Thanks.

I will do what you suggested on v2.

>
> >
> > >
> > > >
> > > > Now that both array and hash maps have support for batched ops in the
> > > > percpu variant, let's provide a convenient macro to declare percpu map
> > > > value types.
> > > >
> > > > Updates the tests to a "reference" usage of the new macro.
> > > >
> > > > Signed-off-by: Pedro Tammela <pctammela@...atatu.com>
> > > > ---
> > > >  tools/lib/bpf/bpf.h                           | 10 ++++
> > > >  tools/testing/selftests/bpf/bpf_util.h        |  7 ---
> > > >  .../bpf/map_tests/htab_map_batch_ops.c        | 48 ++++++++++---------
> > > >  .../selftests/bpf/prog_tests/map_init.c       |  5 +-
> > > >  tools/testing/selftests/bpf/test_maps.c       | 16 ++++---
> > > >  5 files changed, 46 insertions(+), 40 deletions(-)
> > > >
> > >
> > > [...]
> > >
> > > > @@ -400,11 +402,11 @@ static void test_arraymap(unsigned int task, void *data)
> > > >  static void test_arraymap_percpu(unsigned int task, void *data)
> > > >  {
> > > >         unsigned int nr_cpus = bpf_num_possible_cpus();
> > > > -       BPF_DECLARE_PERCPU(long, values);
> > > > +       pcpu_map_value_t values[nr_cpus];
> > > >         int key, next_key, fd, i;
> > > >
> > > >         fd = bpf_create_map(BPF_MAP_TYPE_PERCPU_ARRAY, sizeof(key),
> > > > -                           sizeof(bpf_percpu(values, 0)), 2, 0);
> > > > +                           sizeof(long), 2, 0);
> > > >         if (fd < 0) {
> > > >                 printf("Failed to create arraymap '%s'!\n", strerror(errno));
> > > >                 exit(1);
> > > > @@ -459,7 +461,7 @@ static void test_arraymap_percpu(unsigned int task, void *data)
> > > >  static void test_arraymap_percpu_many_keys(void)
> > > >  {
> > > >         unsigned int nr_cpus = bpf_num_possible_cpus();
> > >
> > > This just sets a bad example for anyone using selftests as an
> > > aspiration for their own code. bpf_num_possible_cpus() does exit(1)
> > > internally if libbpf_num_possible_cpus() returns error. No one should
> > > write real production code like that. So maybe let's provide a better
> > > example instead with error handling and malloc (or perhaps alloca)?
> >
> > OK. Makes sense.
> >
> > >
> > > > -       BPF_DECLARE_PERCPU(long, values);
> > > > +       pcpu_map_value_t values[nr_cpus];
> > > >         /* nr_keys is not too large otherwise the test stresses percpu
> > > >          * allocator more than anything else
> > > >          */
> > > > @@ -467,7 +469,7 @@ static void test_arraymap_percpu_many_keys(void)
> > > >         int key, fd, i;
> > > >
> > > >         fd = bpf_create_map(BPF_MAP_TYPE_PERCPU_ARRAY, sizeof(key),
> > > > -                           sizeof(bpf_percpu(values, 0)), nr_keys, 0);
> > > > +                           sizeof(long), nr_keys, 0);
> > > >         if (fd < 0) {
> > > >                 printf("Failed to create per-cpu arraymap '%s'!\n",
> > > >                        strerror(errno));
> > > > --
> > > > 2.25.1
> > > >

Powered by blists - more mailing lists