Re: [RFC perf,bpf 1/5] perf, bpf: Introduce PERF_RECORD_BPF_EVENT

[Song Liu]

Hi Peter,

> On Nov 8, 2018, at 7:00 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> On Wed, Nov 07, 2018 at 06:25:04PM +0000, Song Liu wrote:
>> 
>> 
>>> On Nov 7, 2018, at 12:40 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>> 
>>> On Tue, Nov 06, 2018 at 12:52:42PM -0800, Song Liu wrote:
>>>> For better performance analysis of BPF programs, this patch introduces
>>>> PERF_RECORD_BPF_EVENT, a new perf_event_type that exposes BPF program
>>>> load/unload information to user space.
>>>> 
>>>>       /*
>>>>        * Record different types of bpf events:
>>>>        *   enum perf_bpf_event_type {
>>>>        *      PERF_BPF_EVENT_UNKNOWN          = 0,
>>>>        *      PERF_BPF_EVENT_PROG_LOAD        = 1,
>>>>        *      PERF_BPF_EVENT_PROG_UNLOAD      = 2,
>>>>        *   };
>>>>        *
>>>>        * struct {
>>>>        *      struct perf_event_header header;
>>>>        *      u16 type;
>>>>        *      u16 flags;
>>>>        *      u32 id;  // prog_id or map_id
>>>>        * };
>>>>        */
>>>>       PERF_RECORD_BPF_EVENT                   = 17,
>>>> 
>>>> PERF_RECORD_BPF_EVENT contains minimal information about the BPF program.
>>>> Perf utility (or other user space tools) should listen to this event and
>>>> fetch more details about the event via BPF syscalls
>>>> (BPF_PROG_GET_FD_BY_ID, BPF_OBJ_GET_INFO_BY_FD, etc.).
>>> 
>>> Why !? You're failing to explain why it cannot provide the full
>>> information there.
>> 
>> Aha, I missed this part. I will add the following to next version. Please
>> let me know if anything is not clear.
> 
>> 
>> This design decision is picked for the following reasons. First, BPF 
>> programs could be loaded-and-jited and/or unloaded before/during/after 
>> perf-record run. Once a BPF programs is unloaded, it is impossible to 
>> recover details of the program. It is impossible to provide the 
>> information through a simple key (like the build ID). Second, BPF prog
>> annotation is under fast developments. Multiple informations will be 
>> added to bpf_prog_info in the next few releases. Including all the
>> information of a BPF program in the perf ring buffer requires frequent 
>> changes to the perf ABI, and thus makes it very difficult to manage 
>> compatibility of perf utility. 
> 
> So I don't agree with that reasoning. If you want symbol information
> you'll just have to commit to some form of ABI. That bpf_prog_info is an
> ABI too.

At the beginning of the perf-record run, perf need to query bpf_prog_info 
of already loaded BPF programs. Therefore, we need to commit to the 
bpf_prog_info ABI. If we also include full information of the BPF program 
in the perf ring buffer, we will commit to TWO ABIs. 

Also, perf-record write the event to perf.data file, so the data need to be 
serialized. This is implemented in patch 4/5. To include the data in the 
ring buffer, we will need another piece of code in the kernel to do the
same serialization work.   

On the other hand, processing BPF load/unload events synchronously should
not introduce too much overhead for meaningful use cases. If many BPF progs
are being loaded/unloaded within short period of time, it is not the steady
state that profiling works care about. 

Would these resolve your concerns? 

Thanks,
Song