Re: [PATCH bpf-next v8 4/9] bpf: Add syscall common attributes support for prog_load

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

On 2026/1/27 10:43, Alexei Starovoitov wrote:
> ,
> 
> 
> On Mon, Jan 26, 2026 at 7:15 AM Leon Hwang <leon.hwang@linux.dev> wrote:
>>
>> The log buffer of common attributes would be confusing with the one in
>> 'union bpf_attr' for BPF_PROG_LOAD.
>>
>> In order to clarify the usage of these two log buffers, they both can be
>> used for logging if:
>>
>> * They are same, including 'log_buf', 'log_level' and 'log_size'.
>> * One of them is missing, then another one will be used for logging.
>>
>> If they both have 'log_buf' but they are not same totally, return -EINVAL.
>>
>> Signed-off-by: Leon Hwang <leon.hwang@linux.dev>
>> ---
>>  include/linux/bpf_verifier.h |  4 +++-
>>  kernel/bpf/log.c             | 29 ++++++++++++++++++++++++++---
>>  kernel/bpf/syscall.c         |  9 ++++++---
>>  3 files changed, 35 insertions(+), 7 deletions(-)
>>
>> diff --git a/include/linux/bpf_verifier.h b/include/linux/bpf_verifier.h
>> index 4a0c5ef296b9..7eb024e83d2d 100644
>> --- a/include/linux/bpf_verifier.h
>> +++ b/include/linux/bpf_verifier.h
>> @@ -637,9 +637,11 @@ struct bpf_log_attr {
>>         u32 log_level;
>>         struct bpf_attrs *attrs;
>>         u32 offsetof_log_true_size;
>> +       struct bpf_attrs *attrs_common;
>>  };
>>
>> -int bpf_prog_load_log_attr_init(struct bpf_log_attr *log_attr, struct bpf_attrs *attrs);
>> +int bpf_prog_load_log_attr_init(struct bpf_log_attr *log_attr, struct bpf_attrs *attrs,
>> +                               struct bpf_attrs *attrs_common);
>>  int bpf_log_attr_finalize(struct bpf_log_attr *log_attr, struct bpf_verifier_log *log);
>>
>>  #define BPF_MAX_SUBPROGS 256
>> diff --git a/kernel/bpf/log.c b/kernel/bpf/log.c
>> index 457b724c4176..c0b816e84384 100644
>> --- a/kernel/bpf/log.c
>> +++ b/kernel/bpf/log.c
>> @@ -865,23 +865,41 @@ void print_insn_state(struct bpf_verifier_env *env, const struct bpf_verifier_st
>>  }
>>
>>  static int bpf_log_attr_init(struct bpf_log_attr *log_attr, struct bpf_attrs *attrs, u64 log_buf,
>> -                            u32 log_size, u32 log_level, int offsetof_log_true_size)
>> +                            u32 log_size, u32 log_level, int offsetof_log_true_size,
>> +                            struct bpf_attrs *attrs_common)
>>  {
>> +       const struct bpf_common_attr *common = attrs_common ? attrs_common->attr : NULL;
>> +
> 
> It seems the only point of struct bpf_attrs is to pass the attrs
> here and remember them in bpf_log_attr.
> It looks unnecessary. Also the cast to void and back to actual
> type doesn't look pretty.
> It seems all of it is to do two copy_to_bpfptr_offset() of log_true_size
> in different places,
> but libbpf will use only one, so let's do only one
> copy_to_bpfptr_offset() and simply the whole thing.
> struct bpf_log_attr only needs
>   u32 offsetof_true_size;
>   bpfptr_t uattr;
> 
> No need to introduce bpf_attrs and bpf_attrs_init() either.
> When bpf_log_attr is inited make the choice of uattr and offset.
> uattr will point at either new bpf_common_attr or old bpf_attr,
> and offsetof_true_size will have 3 possible offsets.
> 
That makes sense.

Then struct bpf_log_attr can be simplified to:

struct bpf_log_attr {
        u32 offsetof_true_size;
        u32 size;
        bpfptr_t uattr;
};

And bpf_log_attr_finalize() would become:

int bpf_log_attr_finalize(struct bpf_log_attr *attr, struct
bpf_verifier_log *log)
{
        ...

        err = bpf_vlog_finalize(log, &log_true_size);

        size = sizeof(log_true_size);
        if (attr->size >= off + size &&
            copy_to_bpfptr_offset(attr->uattr, attr->offsetof_true_size,
&log_true_size, size))
                err = -EFAULT;

        return err;
}

Thanks,
Leon