[RFC PATCH bpf-next] libbpf: make bpf_object__open default to UNSPEC

[RFC PATCH bpf-next] libbpf: make bpf_object__open default to UNSPEC

[Nikita V. Shirokov]

currently by default libbpf's bpf_object__open requires
bpf's program to specify  version in a code because of two things:
1) default prog type is set to KPROBE
2) KPROBE requires (in kernel/bpf/syscall.c) version to be specified

in this RFC i'm proposing change default to UNSPEC and also changing
logic of libbpf that it would reflect what we have today in kernel
(aka only KPROBE type requires for version to be explicitly set).

reason for change:
currently only libbpf requires by default version to be
explicitly set. it would be really hard for mainteiners of other custom
bpf loaders to migrate to libbpf (as they dont control user's code
and migration to the new loader (libbpf) wont be transparent for end
user).

what is going to be broken after this change:
if someone were relying on default to be KPROBE for bpf_object__open
his code will stop to work. however i'm really doubtfull that anyone
is using this for kprobe type of programs (instead of, say, bcc or
other tracing frameworks)

other possible solutions (for discussion, would require more machinery):
add another function like bpf_object__open w/ default to unspec

Signed-off-by: Nikita V. Shirokov <tehnerd@tehnerd.com>
---
 tools/lib/bpf/libbpf.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
index 0f14f7c074c2..ed4212a4c5f9 100644
--- a/tools/lib/bpf/libbpf.c
+++ b/tools/lib/bpf/libbpf.c
@@ -333,7 +333,7 @@ bpf_program__init(void *data, size_t size, char *section_name, int idx,
 	prog->idx = idx;
 	prog->instances.fds = NULL;
 	prog->instances.nr = -1;
-	prog->type = BPF_PROG_TYPE_KPROBE;
+	prog->type = BPF_PROG_TYPE_UNSPEC;
 	prog->btf_fd = -1;
 
 	return 0;
@@ -1649,12 +1649,12 @@ static bool bpf_prog_type__needs_kver(enum bpf_prog_type type)
 	case BPF_PROG_TYPE_LIRC_MODE2:
 	case BPF_PROG_TYPE_SK_REUSEPORT:
 	case BPF_PROG_TYPE_FLOW_DISSECTOR:
-		return false;
 	case BPF_PROG_TYPE_UNSPEC:
-	case BPF_PROG_TYPE_KPROBE:
 	case BPF_PROG_TYPE_TRACEPOINT:
-	case BPF_PROG_TYPE_PERF_EVENT:
 	case BPF_PROG_TYPE_RAW_TRACEPOINT:
+	case BPF_PROG_TYPE_PERF_EVENT:
+		return false;
+	case BPF_PROG_TYPE_KPROBE:
 	default:
 		return true;
 	}
-- 
2.15.1

Re: [RFC PATCH bpf-next] libbpf: make bpf_object__open default to UNSPEC

[Daniel Borkmann]

[ +Wang ]

On 11/22/2018 07:03 AM, Nikita V. Shirokov wrote:
> currently by default libbpf's bpf_object__open requires
> bpf's program to specify  version in a code because of two things:
> 1) default prog type is set to KPROBE
> 2) KPROBE requires (in kernel/bpf/syscall.c) version to be specified
> 
> in this RFC i'm proposing change default to UNSPEC and also changing
> logic of libbpf that it would reflect what we have today in kernel
> (aka only KPROBE type requires for version to be explicitly set).
> 
> reason for change:
> currently only libbpf requires by default version to be
> explicitly set. it would be really hard for mainteiners of other custom
> bpf loaders to migrate to libbpf (as they dont control user's code
> and migration to the new loader (libbpf) wont be transparent for end
> user).
> 
> what is going to be broken after this change:
> if someone were relying on default to be KPROBE for bpf_object__open
> his code will stop to work. however i'm really doubtfull that anyone
> is using this for kprobe type of programs (instead of, say, bcc or
> other tracing frameworks)
> 
> other possible solutions (for discussion, would require more machinery):
> add another function like bpf_object__open w/ default to unspec
> 
> Signed-off-by: Nikita V. Shirokov <tehnerd@tehnerd.com>
> ---
>  tools/lib/bpf/libbpf.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
> index 0f14f7c074c2..ed4212a4c5f9 100644
> --- a/tools/lib/bpf/libbpf.c
> +++ b/tools/lib/bpf/libbpf.c
> @@ -333,7 +333,7 @@ bpf_program__init(void *data, size_t size, char *section_name, int idx,
>  	prog->idx = idx;
>  	prog->instances.fds = NULL;
>  	prog->instances.nr = -1;
> -	prog->type = BPF_PROG_TYPE_KPROBE;
> +	prog->type = BPF_PROG_TYPE_UNSPEC;
>  	prog->btf_fd = -1;

Seems this was mostly for historic reasons, but for a generic library this
would indeed be an odd convention for default. Wang, given 5f44e4c810bf
("tools lib bpf: New API to adjust type of a BPF program"), are you in any
way relying on this default or using things like bpf_program__set_kprobe()
instead which you've added there? If latter, I'd say we should then change
it better now than later when there's even more lib usage (and in particular
before we add official ABI versioning).

>  	return 0;
> @@ -1649,12 +1649,12 @@ static bool bpf_prog_type__needs_kver(enum bpf_prog_type type)
>  	case BPF_PROG_TYPE_LIRC_MODE2:
>  	case BPF_PROG_TYPE_SK_REUSEPORT:
>  	case BPF_PROG_TYPE_FLOW_DISSECTOR:
> -		return false;
>  	case BPF_PROG_TYPE_UNSPEC:
> -	case BPF_PROG_TYPE_KPROBE:
>  	case BPF_PROG_TYPE_TRACEPOINT:
> -	case BPF_PROG_TYPE_PERF_EVENT:
>  	case BPF_PROG_TYPE_RAW_TRACEPOINT:
> +	case BPF_PROG_TYPE_PERF_EVENT:
> +		return false;
> +	case BPF_PROG_TYPE_KPROBE:
>  	default:
>  		return true;
>  	}
> 

Thanks,
Daniel

Re: [RFC PATCH bpf-next] libbpf: make bpf_object__open default to UNSPEC

[Wangnan (F)]

On 2018/11/23 5:52, Daniel Borkmann wrote:
> [ +Wang ]
> 
> On 11/22/2018 07:03 AM, Nikita V. Shirokov wrote:
>> currently by default libbpf's bpf_object__open requires
>> bpf's program to specify  version in a code because of two things:
>> 1) default prog type is set to KPROBE
>> 2) KPROBE requires (in kernel/bpf/syscall.c) version to be specified
>>
>> in this RFC i'm proposing change default to UNSPEC and also changing
>> logic of libbpf that it would reflect what we have today in kernel
>> (aka only KPROBE type requires for version to be explicitly set).
>>
>> reason for change:
>> currently only libbpf requires by default version to be
>> explicitly set. it would be really hard for mainteiners of other custom
>> bpf loaders to migrate to libbpf (as they dont control user's code
>> and migration to the new loader (libbpf) wont be transparent for end
>> user).
>>
>> what is going to be broken after this change:
>> if someone were relying on default to be KPROBE for bpf_object__open
>> his code will stop to work. however i'm really doubtfull that anyone
>> is using this for kprobe type of programs (instead of, say, bcc or
>> other tracing frameworks)
>>
>> other possible solutions (for discussion, would require more machinery):
>> add another function like bpf_object__open w/ default to unspec
>>
>> Signed-off-by: Nikita V. Shirokov <tehnerd@tehnerd.com>
>> ---
>>  tools/lib/bpf/libbpf.c | 8 ++++----
>>  1 file changed, 4 insertions(+), 4 deletions(-)
>>
>> diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
>> index 0f14f7c074c2..ed4212a4c5f9 100644
>> --- a/tools/lib/bpf/libbpf.c
>> +++ b/tools/lib/bpf/libbpf.c
>> @@ -333,7 +333,7 @@ bpf_program__init(void *data, size_t size, char *section_name, int idx,
>>  	prog->idx = idx;
>>  	prog->instances.fds = NULL;
>>  	prog->instances.nr = -1;
>> -	prog->type = BPF_PROG_TYPE_KPROBE;
>> +	prog->type = BPF_PROG_TYPE_UNSPEC;
>>  	prog->btf_fd = -1;
> 
> Seems this was mostly for historic reasons, but for a generic library this
> would indeed be an odd convention for default. Wang, given 5f44e4c810bf
> ("tools lib bpf: New API to adjust type of a BPF program"), are you in any
> way relying on this default or using things like bpf_program__set_kprobe()
> instead which you've added there? If latter, I'd say we should then change
> it better now than later when there's even more lib usage (and in particular
> before we add official ABI versioning).

OK. I don't rely on that now.

Thank you.