Messages in this thread Patch in this message | | | Date | Wed, 20 Jan 2021 13:30:31 -0300 | From | Arnaldo Carvalho de Melo <> | Subject | FIX Re: [PATCH v7 3/3] perf-stat: enable counting events for BPF programs |
| |
Em Wed, Jan 20, 2021 at 10:50:13AM -0300, Arnaldo Carvalho de Melo escreveu: > So sizeof(struct bpf_perf_event_value) == 24 and it is a per-cpu array, the > machine has 24 cpus, why is the kernel thinking it has more and end up zeroing > entries after the 24 cores? Some percpu map subtlety (or obvious thing ;-\) I'm > missing? > > Checking lookups into per cpu maps in sample code now...
(gdb) run stat -b 244 -I 1000 -e cycles Starting program: /root/bin/perf stat -b 244 -I 1000 -e cycles [Thread debugging using libthread_db enabled] Using host libthread_db library "/lib64/libthread_db.so.1". libbpf: elf: skipping unrecognized data section(9) .eh_frame libbpf: elf: skipping relo section(15) .rel.eh_frame for section(9) .eh_frame
Breakpoint 1, bpf_program_profiler__read (evsel=0xce02c0) at util/bpf_counter.c:217 217 if (list_empty(&evsel->bpf_counter_list)) (gdb) p num_ num_cpu num_groups num_leaps num_print_iv num_stmts num_transitions num_warnings_issued num_cpu_bpf num_ifs num_print_interval num_srcfiles num_to_str num_types (gdb) p num_cpu $1 = 24 (gdb) p num_cpu_bpf $2 = 32 (gdb)
Humm, why?
But then libbpf and the sample/bpf/ code use it this way:
diff --git a/tools/perf/util/bpf_counter.c b/tools/perf/util/bpf_counter.c index 8c977f038f497fc1..7dd3d57aba4f620c 100644 --- a/tools/perf/util/bpf_counter.c +++ b/tools/perf/util/bpf_counter.c @@ -207,7 +207,8 @@ static int bpf_program_profiler__enable(struct evsel *evsel) static int bpf_program_profiler__read(struct evsel *evsel) { int num_cpu = evsel__nr_cpus(evsel); - struct bpf_perf_event_value values[num_cpu]; + int num_cpu_bpf = libbpf_num_possible_cpus(); + struct bpf_perf_event_value values[num_cpu > num_cpu_bpf ? num_cpu : num_cpu_bpf]; struct bpf_counter *counter; int reading_map_fd; __u32 key = 0; ------------------------------------------------------------- [root@five ~]# cat /sys/devices/system/cpu/possible 0-31 [root@five ~]#
I bet that in your test systems evsel__nr_cpus(evsel) matches /sys/devices/system/cpu/possible and thus you don't see the problem.
evsel__nr_cpus(evsel) uses what is in:
[acme@five perf]$ cat /sys/devices/system/cpu/online 0-23 [acme@five perf]$
So that is the reason for the problem and the fix is to use libbpf_num_possible_cpus(), I'll bolt that into the patch that introduced that code.
- Arnaldo
| |