Re: [PATCH bpf-next] bpf: Generalize data copying for percpu maps

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

On 4/9/25 07:39, Andrii Nakryiko wrote:
> On Wed, Sep 3, 2025 at 10:36 AM Alexei Starovoitov
> <alexei.starovoitov@gmail.com> wrote:
>>
>> On Wed, Sep 3, 2025 at 10:04 AM Leon Hwang <leon.hwang@linux.dev> wrote:
>>>
>>> While adding support for the BPF_F_CPU and BPF_F_ALL_CPUS flags, the data
>>> copying logic of the following percpu map types needs to be updated:
>>>
>>> * percpu_array
>>> * percpu_hash
>>> * lru_percpu_hash
>>> * percpu_cgroup_storage
>>>
>>> Following Andrii’s suggestion[0], this patch refactors the data copying
> 
> as flattering as that is, "Andrii's suggestion" is no justification
> why the patch is correct :)
> 

:)

>>> logic by introducing two helpers:
>>>
>>> * `bpf_percpu_copy_to_user()`
>>> * `bpf_percpu_copy_from_user()`
>>>
>>> This prepares the codebase for the upcoming CPU flag support.
>>>
>>> [0] https://lore.kernel.org/bpf/20250827164509.7401-1-leon.hwang@linux.dev/
>>>
>>> Suggested-by: Andrii Nakryiko <andrii.nakryiko@gmail.com>
>>> Signed-off-by: Leon Hwang <leon.hwang@linux.dev>
>>> ---
>>>  include/linux/bpf.h        | 29 ++++++++++++++++++++++++++++-
>>>  kernel/bpf/arraymap.c      | 14 ++------------
>>>  kernel/bpf/hashtab.c       | 20 +++-----------------
>>>  kernel/bpf/local_storage.c | 18 ++++++------------
>>>  4 files changed, 39 insertions(+), 42 deletions(-)
>>>
>>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>>> index 8f6e87f0f3a89..2dc0299a2da50 100644
>>> --- a/include/linux/bpf.h
>>> +++ b/include/linux/bpf.h
>>> @@ -547,6 +547,34 @@ static inline void copy_map_value_long(struct bpf_map *map, void *dst, void *src
>>>         bpf_obj_memcpy(map->record, dst, src, map->value_size, true);
>>>  }
>>>
>>> +#ifdef CONFIG_BPF_SYSCALL
>>> +static inline void bpf_percpu_copy_to_user(struct bpf_map *map, void __percpu *pptr, void *value,
>>> +                                          u32 size)
>>> +{
>>> +       int cpu, off = 0;
>>> +
>>> +       for_each_possible_cpu(cpu) {
>>> +               copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
>>> +               check_and_init_map_value(map, value + off);
> 
> I still maintain that this makes zero sense... value+off is memory
> that we'll copy_to_user, why are we setting refcount to 1, or
> rb_node/list_node to "proper empty node" is absolutely not clear... it
> feels like we can drop check_and_init_map_value() altogether and be
> absolutely no worse. If anything, memset(0) would be nicer, but I
> guess we didn't have it to begin with, so no need to add it now.
> 

Agreed.

As 'copy_map_value_long()' won't copy those fields,
'check_and_init_map_value()' is unnecessary here.

>>> +               off += size;
>>> +       }
>>> +}
>>> +
>>> +void bpf_obj_free_fields(const struct btf_record *rec, void *obj);
>>> +
>>> +static inline void bpf_percpu_copy_from_user(struct bpf_map *map, void __percpu *pptr, void *value,
>>> +                                            u32 size)
>>> +{
>>> +       int cpu, off = 0;
>>> +
>>> +       for_each_possible_cpu(cpu) {
>>> +               copy_map_value_long(map, per_cpu_ptr(pptr, cpu), value + off);
> 
> copy_map_value_long is generalization of bpf_long_memcpy, and so it
> would be good to call this out to explain why your refactoring is
> correct
> 

No.

It shouldn't call bpf_long_memcpy() before bpf_obj_free_fields(), or it
will overwrite those fields data used for bpf_obj_free_fields().

It would be better to call bpf_obj_free_fields() then bpf_long_memcpy().

>>> +               bpf_obj_free_fields(map->record, per_cpu_ptr(pptr, cpu));
>>> +               off += size;
>>> +       }
>>> +}
>>> +#endif
>>> +
>>>  static inline void bpf_obj_swap_uptrs(const struct btf_record *rec, void *dst, void *src)
>>>  {
>>>         unsigned long *src_uptr, *dst_uptr;
>>> @@ -2417,7 +2445,6 @@ struct btf_record *btf_record_dup(const struct btf_record *rec);
>>>  bool btf_record_equal(const struct btf_record *rec_a, const struct btf_record *rec_b);
>>>  void bpf_obj_free_timer(const struct btf_record *rec, void *obj);
>>>  void bpf_obj_free_workqueue(const struct btf_record *rec, void *obj);
>>> -void bpf_obj_free_fields(const struct btf_record *rec, void *obj);
>>>  void __bpf_obj_drop_impl(void *p, const struct btf_record *rec, bool percpu);
>>>
>>>  struct bpf_map *bpf_map_get(u32 ufd);
>>> diff --git a/kernel/bpf/arraymap.c b/kernel/bpf/arraymap.c
>>> index 3d080916faf97..6be9c54604503 100644
>>> --- a/kernel/bpf/arraymap.c
>>> +++ b/kernel/bpf/arraymap.c
>>> @@ -300,7 +300,6 @@ int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value)
>>>         struct bpf_array *array = container_of(map, struct bpf_array, map);
>>>         u32 index = *(u32 *)key;
>>>         void __percpu *pptr;
>>> -       int cpu, off = 0;
>>>         u32 size;
>>>
>>>         if (unlikely(index >= array->map.max_entries))
>>> @@ -313,11 +312,7 @@ int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value)
>>>         size = array->elem_size;
>>>         rcu_read_lock();
>>>         pptr = array->pptrs[index & array->index_mask];
>>> -       for_each_possible_cpu(cpu) {
>>> -               copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
>>> -               check_and_init_map_value(map, value + off);
>>> -               off += size;
>>> -       }
>>> +       bpf_percpu_copy_to_user(map, pptr, value, size);
>>>         rcu_read_unlock();
>>>         return 0;
>>>  }
>>> @@ -387,7 +382,6 @@ int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value,
>>>         struct bpf_array *array = container_of(map, struct bpf_array, map);
>>>         u32 index = *(u32 *)key;
>>>         void __percpu *pptr;
>>> -       int cpu, off = 0;
>>>         u32 size;
>>>
>>>         if (unlikely(map_flags > BPF_EXIST))
>>> @@ -411,11 +405,7 @@ int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value,
>>>         size = array->elem_size;
>>>         rcu_read_lock();
>>>         pptr = array->pptrs[index & array->index_mask];
>>> -       for_each_possible_cpu(cpu) {
>>> -               copy_map_value_long(map, per_cpu_ptr(pptr, cpu), value + off);
>>> -               bpf_obj_free_fields(array->map.record, per_cpu_ptr(pptr, cpu));
>>> -               off += size;
>>> -       }
>>> +       bpf_percpu_copy_from_user(map, pptr, value, size);
>>>         rcu_read_unlock();
>>>         return 0;
>>>  }
>>> diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c
>>> index 71f9931ac64cd..5f0f3c00dbb74 100644
>>> --- a/kernel/bpf/hashtab.c
>>> +++ b/kernel/bpf/hashtab.c
>>> @@ -944,12 +944,8 @@ static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
>>>                 copy_map_value(&htab->map, this_cpu_ptr(pptr), value);
>>>         } else {
>>>                 u32 size = round_up(htab->map.value_size, 8);
>>> -               int off = 0, cpu;
>>>
>>> -               for_each_possible_cpu(cpu) {
>>> -                       copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value + off);
>>> -                       off += size;
>>> -               }
>>> +               bpf_percpu_copy_from_user(&htab->map, pptr, value, size);
>>
>> This is not a refactor. There is a significant change in the logic.
>> Why is it needed? Bug fix or introducing a bug?
> 
> this is preparation for that BPF_F_CPU/BPF_F_ALLCPUS, but I agree that
> it would be better to include as preparatory patch in the actual patch
> set
> 

Ack.

I'll move this patch into the patch set of BPF_F_CPU/BPF_F_ALLCPUS flags.

>>
>> The names to_user and from_user are wrong.
>> There is no user space memory involved.
> 
> This was my suggestion because we either are copying user-supplied
> data or copying data back to user. Strictly speaking it's all kernel
> memory (copy_from_user/copy_to_user is done afterwards by the caller),
> but that's the intent.
> 
> Maybe "copy_in" and "copy_out" would be better, I don't know. But
> there is certainly a direction here w.r.t. user space provided data
> (note, this is not BPF program-side logic).
> 

'bpf_percpu_copy_data()' and 'bpf_percpu_update_data()' would be better,
as "copy_data" is used in those 'bpf_percpu_*_copy()' functions and
"update_data" is used in those 'bpf_percpu_*_update()' functions.

Thanks,
Leon