On 2/5/19 6:26 PM, Alexei Starovoitov wrote:
> On Tue, Feb 05, 2019 at 11:48:23AM -0800, Yonghong Song wrote:
>> The kernel verifier has three levels of logs:
>> 0: no logs
>> 1: logs mostly useful
>> > 1: verbose
>>
>> Current libbpf API functions bpf_load_program_xattr() and
>> bpf_load_program() cannot specify log_level.
>> The bcc, however, provides an interface for user to
>> specify log_level 2 for verbose output.
>>
>> This patch added log_level into structure
>> bpf_load_program_attr, so users, including bcc, can use
>> bpf_load_program_xattr() to change log_level.
>>
>> The bpf selftest test_sock.c is modified to enable log_level = 2.
>> If the "verbose" in test_sock.c is changed to true,
>> the test will output logs like below:
>> $ ./test_sock
>> func#0 @0
>> 0: R1=ctx(id=0,off=0,imm=0) R10=fp0,call_-1
>> 0: (bf) r6 = r1
>> 1: R1=ctx(id=0,off=0,imm=0) R6_w=ctx(id=0,off=0,imm=0) R10=fp0,call_-1
>> 1: (61) r7 = *(u32 *)(r6 +28)
>> invalid bpf_context access off=28 size=4
>>
>> Test case: bind4 load with invalid access: src_ip6 .. [PASS]
>> ...
>> Test case: bind6 allow all .. [PASS]
>> Summary: 16 PASSED, 0 FAILED
>>
>> Some test_sock tests are negative tests and verbose verifier
>> log will be printed out as shown in the above.
>>
>> Signed-off-by: Yonghong Song <y...@fb.com>
>> ---
>> tools/lib/bpf/bpf.c | 4 +++-
>> tools/lib/bpf/bpf.h | 1 +
>> tools/testing/selftests/bpf/test_sock.c | 9 ++++++++-
>> 3 files changed, 12 insertions(+), 2 deletions(-)
>>
>> diff --git a/tools/lib/bpf/bpf.c b/tools/lib/bpf/bpf.c
>> index 3defad77dc7a..78aa8c2b1a5c 100644
>> --- a/tools/lib/bpf/bpf.c
>> +++ b/tools/lib/bpf/bpf.c
>> @@ -214,6 +214,7 @@ int bpf_load_program_xattr(const struct
>> bpf_load_program_attr *load_attr,
>> {
>> void *finfo = NULL, *linfo = NULL;
>> union bpf_attr attr;
>> + __u32 log_level;
>> __u32 name_len;
>> int fd;
>>
>> @@ -292,7 +293,8 @@ int bpf_load_program_xattr(const struct
>> bpf_load_program_attr *load_attr,
>> /* Try again with log */
>> attr.log_buf = ptr_to_u64(log_buf);
>> attr.log_size = log_buf_sz;
>> - attr.log_level = 1;
>> + log_level = load_attr->log_level;
>> + attr.log_level = (log_level >= 2) ? log_level : 1;
>> log_buf[0] = 0;
>> fd = sys_bpf_prog_load(&attr, sizeof(attr));
>> done:
>> diff --git a/tools/lib/bpf/bpf.h b/tools/lib/bpf/bpf.h
>> index ed09eed2dc3b..15a8e22e8eae 100644
>> --- a/tools/lib/bpf/bpf.h
>> +++ b/tools/lib/bpf/bpf.h
>> @@ -76,6 +76,7 @@ struct bpf_load_program_attr {
>> const struct bpf_insn *insns;
>> size_t insns_cnt;
>> const char *license;
>> + __u32 log_level;
>> __u32 kern_version;
>> __u32 prog_ifindex;
>> __u32 prog_btf_fd;
>
> this will break binary compatibility in libbpf api.
> Please add new fields always to the end of *_attr structs.
I felt that we were still at 0.0.x stage and still could tweak
the api a little bit so I used the above sequence to
mimic the kernel prog_load attr sequence. I do agree that
in general we should add to the end of data structure.
I can change to the end of structure we still decided
if exposing log_level=2 is needed.
>
> Also why not to silence bcc instead?
> Let it treat log_level > 1 as log_level=1
> I don't think anyone but the most extreme verifier hacker used level=2.
In general, that is true. I just used the feature a couple of
weeks ago to toubleshoot a verifier error... But most cases
log_level=1 should be sufficient.
> Personally I don't remember when was the last time I used it.
> It seem like a niche feature that we can safely remove in bcc.
Will discuss a little more in bcc community and then make
a decision.
