Re: [PATCH bpf-next v2 02/20] bpf: allow loading of a bpf_iter program

[Yonghong Song <yhs@fb.com>]

On 5/5/20 2:29 PM, Andrii Nakryiko wrote:
> On Sun, May 3, 2020 at 11:26 PM Yonghong Song <yhs@fb.com> wrote:
>>
>> A bpf_iter program is a tracing program with attach type
>> BPF_TRACE_ITER. The load attribute
>>    attach_btf_id
>> is used by the verifier against a particular kernel function,
>> which represents a target, e.g., __bpf_iter__bpf_map
>> for target bpf_map which is implemented later.
>>
>> The program return value must be 0 or 1 for now.
>>    0 : successful, except potential seq_file buffer overflow
>>        which is handled by seq_file reader.
>>    1 : request to restart the same object
> 
> This bit is interesting. Is the idea that if BPF program also wants to
> send something over, say, perf_buffer, but fails, it can "request"
> same execution again? I wonder if typical libc fread() implementation

Yes. The bpf_seq_read() can handle this the same as any other
retry request. The following is current mapping.
    bpf program return 0   ---> seq_ops->show() return 0
    bpf program return 1   ---> seq_ops->show() return -EAGAIN

> would handle EAGAIN properly, it seems more driven towards
> non-blocking I/O?

I did not have a test for this in current patch set for bpf program
returning 1. Will add a test in the next version.

> 
> On the other hand, following start/show/next logic for seq_file
> iteration, requesting skipping element seems useful. It would allow
> (in some cases) to "speculatively" generate output and at some point
> realize that this is not an element we actually want in the output and
> request to ignore that output.
> 
> Don't know how useful the latter is going to be in practice, but just
> something to keep in mind for the future, I guess...
> 
>>
>> In the future, other return values may be used for filtering or
>> teminating the iterator.
>>
>> Signed-off-by: Yonghong Song <yhs@fb.com>
>> ---
>>   include/linux/bpf.h            |  3 +++
>>   include/uapi/linux/bpf.h       |  1 +
>>   kernel/bpf/bpf_iter.c          | 30 ++++++++++++++++++++++++++++++
>>   kernel/bpf/verifier.c          | 21 +++++++++++++++++++++
>>   tools/include/uapi/linux/bpf.h |  1 +
>>   5 files changed, 56 insertions(+)
>>
> 
> [...]
> 
> 
>> +
>> +bool bpf_iter_prog_supported(struct bpf_prog *prog)
>> +{
>> +       const char *attach_fname = prog->aux->attach_func_name;
>> +       u32 prog_btf_id = prog->aux->attach_btf_id;
>> +       const char *prefix = BPF_ITER_FUNC_PREFIX;
>> +       struct bpf_iter_target_info *tinfo;
>> +       int prefix_len = strlen(prefix);
>> +       bool supported = false;
>> +
>> +       if (strncmp(attach_fname, prefix, prefix_len))
>> +               return false;
>> +
>> +       mutex_lock(&targets_mutex);
>> +       list_for_each_entry(tinfo, &targets, list) {
>> +               if (tinfo->btf_id && tinfo->btf_id == prog_btf_id) {
>> +                       supported = true;
>> +                       break;
>> +               }
>> +               if (!strcmp(attach_fname + prefix_len, tinfo->target)) {
>> +                       tinfo->btf_id = prog->aux->attach_btf_id;
> 
> This target_info->btf_id caching here is a bit subtle and easy to
> miss, it would be nice to have a code calling this out explicitly.

Will do.

> Thanks!
> 
>> +                       supported = true;
>> +                       break;
>> +               }
>> +       }
>> +       mutex_unlock(&targets_mutex);
>> +
>> +       return supported;
>> +}
>> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
>> index 70ad009577f8..d725ff7d11db 100644
>> --- a/kernel/bpf/verifier.c
>> +++ b/kernel/bpf/verifier.c
>> @@ -7101,6 +7101,10 @@ static int check_return_code(struct bpf_verifier_env *env)
>>                          return 0;
>>                  range = tnum_const(0);
>>                  break;
>> +       case BPF_PROG_TYPE_TRACING:
>> +               if (env->prog->expected_attach_type != BPF_TRACE_ITER)
>> +                       return 0;
> 
> Commit message mentions enforcing [0, 1], shouldn't it be done here?

The default range is [0, 1], hence no explicit assignment here.

static int check_return_code(struct bpf_verifier_env *env)
{
         struct tnum enforce_attach_type_range = tnum_unknown;
         const struct bpf_prog *prog = env->prog;
         struct bpf_reg_state *reg;
         struct tnum range = tnum_range(0, 1);
......

> 
> 
>> +               break;
>>          default:
>>                  return 0;
>>          }
> 
> [...]
>