Re: [RFC PATCH bpf-next 2/3] bpf, libbpf: Support BPF_F_CPU for percpu_array map

[Leon Hwang <leon.hwang@linux.dev>]

On 2025/7/3 01:30, Andrii Nakryiko wrote:
> On Wed, Jul 2, 2025 at 10:28 AM Leon Hwang <leon.hwang@linux.dev> wrote:
>>
>>
>>
>> On 2025/7/2 04:22, Andrii Nakryiko wrote:
>>> On Tue, Jun 24, 2025 at 9:55 AM Leon Hwang <leon.hwang@linux.dev> wrote:
>>>>
>>>> This patch adds libbpf support for the BPF_F_CPU flag in percpu_array maps,
>>>> introducing the following APIs:
>>>>
>>>> 1. bpf_map_update_elem_opts(): update with struct bpf_map_update_elem_opts
>>>> 2. bpf_map_lookup_elem_opts(): lookup with struct bpf_map_lookup_elem_opts
>>>> 3. bpf_map__update_elem_opts(): high-level wrapper with input validation
>>>> 4. bpf_map__lookup_elem_opts(): high-level wrapper with input validation
>>>>
>>>> Behavior:
>>>>
>>>> * If opts->cpu == 0xFFFFFFFF, the update is applied to all CPUs.
>>>> * Otherwise, it applies only to the specified CPU.
>>>> * Lookup APIs retrieve values from the target CPU when BPF_F_CPU is used.
>>>>
>>>> Signed-off-by: Leon Hwang <leon.hwang@linux.dev>
>>>> ---
>>>>  tools/lib/bpf/bpf.c           | 37 +++++++++++++++++++++++
>>>>  tools/lib/bpf/bpf.h           | 35 +++++++++++++++++++++-
>>>>  tools/lib/bpf/libbpf.c        | 56 +++++++++++++++++++++++++++++++++++
>>>>  tools/lib/bpf/libbpf.h        | 45 ++++++++++++++++++++++++++++
>>>>  tools/lib/bpf/libbpf.map      |  4 +++
>>>>  tools/lib/bpf/libbpf_common.h | 12 ++++++++
>>>>  6 files changed, 188 insertions(+), 1 deletion(-)
>>>>
> 
> [...]
> 
>>>>  };
>>>> -#define bpf_map_batch_opts__last_field flags
>>>> +#define bpf_map_batch_opts__last_field cpu
>>>>
>>>>
>>>>  /**
>>>> @@ -286,6 +315,10 @@ LIBBPF_API int bpf_map_lookup_and_delete_batch(int fd, void *in_batch,
>>>>   *    Update spin_lock-ed map elements. This must be
>>>>   *    specified if the map value contains a spinlock.
>>>>   *
>>>> + * **BPF_F_CPU**
>>>> + *    As for percpu map, update value on all CPUs if **opts->cpu** is
>>>> + *    0xFFFFFFFF, or on specified CPU otherwise.
>>>> + *
>>>>   * @param fd BPF map file descriptor
>>>>   * @param keys pointer to an array of *count* keys
>>>>   * @param values pointer to an array of *count* values
>>>> diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
>>>> index 6445165a24f2..30400bdc20d9 100644
>>>> --- a/tools/lib/bpf/libbpf.c
>>>> +++ b/tools/lib/bpf/libbpf.c
>>>> @@ -10636,6 +10636,34 @@ int bpf_map__lookup_elem(const struct bpf_map *map,
>>>>         return bpf_map_lookup_elem_flags(map->fd, key, value, flags);
>>>>  }
>>>>
>>>> +int bpf_map__lookup_elem_opts(const struct bpf_map *map, const void *key,
>>>> +                             size_t key_sz, void *value, size_t value_sz,
>>>> +                             const struct bpf_map_lookup_elem_opts *opts)
>>>> +{
>>>> +       int nr_cpus = libbpf_num_possible_cpus();
>>>> +       __u32 cpu = OPTS_GET(opts, cpu, nr_cpus);
>>>> +       __u64 flags = OPTS_GET(opts, flags, 0);
>>>> +       int err;
>>>> +
>>>> +       if (flags & BPF_F_CPU) {
>>>> +               if (map->def.type != BPF_MAP_TYPE_PERCPU_ARRAY)
>>>> +                       return -EINVAL;
>>>> +               if (cpu >= nr_cpus)
>>>> +                       return -E2BIG;
>>>> +               if (map->def.value_size != value_sz) {
>>>> +                       pr_warn("map '%s': unexpected value size %zu provided, expected %u\n",
>>>> +                               map->name, value_sz, map->def.value_size);
>>>> +                       return -EINVAL;
>>>> +               }
>>>
>>> shouldn't this go into validate_map_op?..
>>>
>>
>> It should.
>>
>> However, to avoid making validate_map_op really complicated, I'd like to
>> add validate_map_cpu_op to wrap checking cpu and validate_map_op.
> 
> validate_map_op is meant to handle all the different conditions, let's
> keep all that in one function
> 

Got it.

Thanks,
Leon