Re: [PATCH v3 perf, bpf-next 1/4] perf, bpf: Introduce PERF_RECORD_BPF_EVENT

[Song Liu]

> On Dec 13, 2018, at 7:25 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> On Wed, Dec 12, 2018 at 06:56:11PM +0000, Song Liu wrote:
>> 
>> 
>>> On Dec 12, 2018, at 10:05 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>> 
>>> On Wed, Dec 12, 2018 at 05:09:17PM +0000, Song Liu wrote:
>>>>> And while this tracks the bpf kallsyms, it does not do all kallsyms.
>>>>> 
>>>>> .... Oooh, I see the problem, everybody is doing their own custom
>>>>> kallsym_{add,del}() thing, instead of having that in generic code :-(
>>>>> 
>>>>> This, for example, doesn't track module load/unload nor ftrace
>>>>> trampolines, even though both affect kallsyms.
>>>> 
>>>> I think we can use PERF_RECORD_MMAP(or MMAP2) for module load/unload. 
>>>> That could be separate sets of patches. 
>>> 
>>> So I would actually like to move bpf_lock/bpf_kallsyms/bpf_tree +
>>> bpf_prog_kallsyms_*() + __bpf_address_lookup() into kernel/kallsyms.c
>>> and also have ftrace use that.
>>> 
>>> Because currently the ftrace stuff is otherwise invisible.
>>> 
>>> A generic kallsym register/unregister for any JIT.
>> 
>> I guess this is _not_ a requirement for this patchset? BPF program has
>> special data (id, sub_id, tag) that we need PERF_RECORD_BPF_EVENT. So 
>> this patchset should be orthogonal to the generic kallsym framework?
> 
> Well, it is a question of ABI. I don't like mixing the kallsym updates
> with the BPF updates.

I have been always thinking the two is one update: "mapping this BPF 
program to this ksym". 

On the other hand, if we really want to separate the two. I guess we 
need two PERF_RECORD_*:

       /*
        * PERF_RECORD_KSYM_ADD/DEL or MMAP3/MUNMAP3
        *
        * struct {
        *      struct perf_event_header header;
        *      u64                             addr;
        *      u64                             len;
        *      char                            name[];
        *      struct sample_id                sample_id;
        * };
        */

and

       /*
        * PERF_RECORD_BPF_EVENT
        *
        * struct {
        *      struct perf_event_header header;
        *      u32                             type;
        *      u32                             flags;
        *      u32                             id; // prog_id or other id
        *      u32                             sub_id; // subprog id
        *
        *      // for bpf_prog types, bpf prog or subprog
        *      u8                              tag[BPF_TAG_SIZE];
        *      struct sample_id                sample_id;
        * };
        */

In this case, PERF_RECORD_BPF_EVENT is only needed when user want
annotation. When annotation is needed, kernel will generate both
record for each BPF program load/unload. Then, user space will do 
some work to match the two. 

Personally, I think this is not as clean as current version. But
it would work. 

Would you recommend we go on this direction? 

Thanks,
Song