Re: [PATCH bpf-next 2/3] bpf: emit RECORD_MMAP events for bpf prog load/unload

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Wed, Oct 17, 2018 at 09:11:40AM -0300, Arnaldo Carvalho de Melo escreveu:
> Adding Alexey, Jiri and Namhyung as they worked/are working on
> multithreading 'perf record'.
> 
> Em Tue, Oct 16, 2018 at 11:43:11PM -0700, Song Liu escreveu:
> > On Tue, Oct 16, 2018 at 4:43 PM David Ahern <dsahern@gmail.com> wrote:
> > > On 10/15/18 4:33 PM, Song Liu wrote:
> > > > I am working with Alexei on the idea of fetching BPF program information via
> > > > BPF_OBJ_GET_INFO_BY_FD cmd. I added PERF_RECORD_BPF_EVENT
> > > > to perf_event_type, and dumped these events to perf event ring buffer.
> 
> > > > I found that perf will not process event until the end of perf-record:
> 
> > > > root@virt-test:~# ~/perf record -ag -- sleep 10
> > > > ...... 10 seconds later
> > > > [ perf record: Woken up 34 times to write data ]
> > > > machine__process_bpf_event: prog_id 6 loaded
> > > > machine__process_bpf_event: prog_id 6 unloaded
> > > > [ perf record: Captured and wrote 9.337 MB perf.data (93178 samples) ]
> 
> > > > In this example, the bpf program was loaded and then unloaded in
> > > > another terminal. When machine__process_bpf_event() processes
> > > > the load event, the bpf program is already unloaded. Therefore,
> > > > machine__process_bpf_event() will not be able to get information
> > > > about the program via BPF_OBJ_GET_INFO_BY_FD cmd.
> 
> > > > To solve this problem, we will need to run BPF_OBJ_GET_INFO_BY_FD
> > > > as soon as perf get the event from kernel. I looked around the perf
> > > > code for a while. But I haven't found a good example where some
> > > > events are processed before the end of perf-record. Could you
> > > > please help me with this?
> 
> > > perf record does not process events as they are generated. Its sole job
> > > is pushing data from the maps to a file as fast as possible meaning in
> > > bulk based on current read and write locations.
> 
> > > Adding code to process events will add significant overhead to the
> > > record command and will not really solve your race problem.
> 
> > I agree that processing events while recording has significant overhead.
> > In this case, perf user space need to know details about the the jited BPF
> > program. It is impossible to pass all these details to user space through
> > the relatively stable ring_buffer API. Therefore, some processing of the
> > data is necessary (get bpf prog_id from ring buffer, and then fetch program
> > details via BPF_OBJ_GET_INFO_BY_FD.
>  
> > I have some idea on processing important data with relatively low overhead.
> > Let me try implement it.
> 
> Well, you could have a separate thread processing just those kinds of
> events, associate it with a dummy event where you only ask for
> PERF_RECORD_BPF_EVENTs.
> 
> Here is how to setup the PERF_TYPE_SOFTWARE/PERF_COUNT_SW_DUMMY
> perf_event_attr:
> 
> [root@seventh ~]# perf record -vv -e dummy sleep 01
> ------------------------------------------------------------
> perf_event_attr:
>   type                             1
>   size                             112
>   config                           0x9
>   { sample_period, sample_freq }   4000
>   sample_type                      IP|TID|TIME|PERIOD
>   disabled                         1
>   inherit                          1

These you would have disabled, no need for
PERF_RECORD_{MMAP*,COMM,FORK,EXIT} just PERF_RECORD_BPF_EVENT

>   mmap                             1
>   comm                             1
>   task                             1
>   mmap2                            1
>   comm_exec                        1


>   freq                             1
>   enable_on_exec                   1
>   sample_id_all                    1
>   exclude_guest                    1
> ------------------------------------------------------------
> sys_perf_event_open: pid 12046  cpu 0  group_fd -1  flags 0x8 = 4
> sys_perf_event_open: pid 12046  cpu 1  group_fd -1  flags 0x8 = 5
> sys_perf_event_open: pid 12046  cpu 2  group_fd -1  flags 0x8 = 6
> sys_perf_event_open: pid 12046  cpu 3  group_fd -1  flags 0x8 = 8
> mmap size 528384B
> perf event ring buffer mmapped per cpu
> Synthesizing TSC conversion information
> [ perf record: Woken up 1 times to write data ]
> [ perf record: Captured and wrote 0.014 MB perf.data ]
> [root@seventh ~]#
> 
> [root@seventh ~]# perf evlist -v
> dummy: type: 1, size: 112, config: 0x9, { sample_period, sample_freq }: 4000, sample_type: IP|TID|TIME|PERIOD, disabled: 1, inherit: 1, mmap: 1, comm: 1, freq: 1, enable_on_exec: 1, task: 1, sample_id_all: 1, exclude_guest: 1, mmap2: 1, comm_exec: 1
> [root@seventh ~]# 
> 
> There is work ongoing in dumping one file per cpu and then, at post
> processing time merging all those files to get ordering, so one more
> file, for these VIP events, that require per-event processing would be
> ordered at that time with all the other per-cpu files.
> 
> - Arnaldo