Re: [PATCH v7 3/3] perf-stat: enable counting events for BPF programs

From: Arnaldo Carvalho de Melo
Date: Tue Jan 19 2021 - 17:32:10 EST

Em Tue, Jan 19, 2021 at 09:54:50PM +0000, Song Liu escreveu:
>
>
> > On Jan 19, 2021, at 8:31 AM, Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:
> >
> > Em Tue, Jan 19, 2021 at 11:42:49AM -0300, Arnaldo Carvalho de Melo escreveu:
> >> Em Tue, Jan 19, 2021 at 11:31:44AM -0300, Arnaldo Carvalho de Melo escreveu:
> >>> Em Tue, Jan 19, 2021 at 12:48:19AM +0000, Song Liu escreveu:
> >>>>> On Jan 18, 2021, at 11:38 AM, Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:
> >>>> We are looking at two issues:
> >>>> 1. Cannot recursively attach;
> >>>> 2. prog FD 3 doesn't have valid btf.
> >>
> >>>> #1 was caused by the verifier disallowing attaching fentry/fexit program
> >>>> to program with type BPF_PROG_TYPE_TRACING (in bpf_check_attach_target).
> >>>> This constraint was added when we only had fentry/fexit in the TRACING
> >>>> type. We have extended the TRACING type to many other use cases, like
> >>>> "tp_btf/", "fmod_ret" and "iter/". Therefore, it is good time to revisit
> >>>> this constraint. I will work on this.
> >>
> >>>> For #2, we require the target program to have BTF. I guess we won't remove
> >>>> this requirement.
> >>
> >>>> While I work on improving #1, could you please test with some kprobe
> >>>> programs? For example, we can use fileslower.py from bcc.
> >>
> >>> Sure, and please consider improving the error messages to state what you
> >>> described above.
> >>
> >> Terminal 1:
> >>
> >> [root@five perf]# perf trace -e 5sec.c
> >> ^C
> >> [root@five perf]# cat 5sec.c
> >> #include <bpf.h>
> >>
> >> #define NSEC_PER_SEC 1000000000L
> >>
> >> int probe(hrtimer_nanosleep, rqtp)(void *ctx, int err, long long sec)
> >> {
> >> return sec / NSEC_PER_SEC == 5;
> >> }
> >>
> >> license(GPL);
> >> [root@five perf]# perf trace -e 5sec.c/max-stack=16/
> >> 0.000 sleep/3739435 perf_bpf_probe:hrtimer_nanosleep(__probe_ip: -1743337312, rqtp: 5000000000)
> >> hrtimer_nanosleep ([kernel.kallsyms])
> >> common_nsleep ([kernel.kallsyms])
> >> __x64_sys_clock_nanosleep ([kernel.kallsyms])
> >> do_syscall_64 ([kernel.kallsyms])
> >> entry_SYSCALL_64_after_hwframe ([kernel.kallsyms])
> >> __clock_nanosleep_2 (/usr/lib64/libc-2.32.so)
> >>
> >>
> >> Terminal 2:
> >>
> >> [root@five ~]# perf stat -e cycles -b 180 -I 1000
> >> libbpf: elf: skipping unrecognized data section(9) .eh_frame
> >> libbpf: elf: skipping relo section(15) .rel.eh_frame for section(9) .eh_frame
> >> perf: util/bpf_counter.c:227: bpf_program_profiler__read: Assertion `skel != NULL' failed.
> >> Aborted (core dumped)
> >
> > Out to lunch, will continue later, but this may help you figure this out
> > till then :)
> >
> > Starting program: /root/bin/perf stat -e cycles -b 244 -I 1000
> > [Thread debugging using libthread_db enabled]
> > Using host libthread_db library "/lib64/libthread_db.so.1".
> >
> > Breakpoint 1, bpf_program_profiler_load_one (evsel=0xce02c0, prog_id=244) at util/bpf_counter.c:96
> > 96 {
> > (gdb) n
> > 104 prog_fd = bpf_prog_get_fd_by_id(prog_id);
> > (gdb)
> > 105 if (prog_fd < 0) {
> > (gdb)
> > 109 counter = bpf_counter_alloc();
> > (gdb)
> > 110 if (!counter) {
> > (gdb) n
> > 115 skel = bpf_prog_profiler_bpf__open();
> > (gdb) p counter
> > $9 = (struct bpf_counter *) 0xce09e0
> > (gdb) p *counter
> > $10 = {skel = 0x0, list = {next = 0xce09e8, prev = 0xce09e8}}
> > (gdb) p *counter
> > $11 = {skel = 0x0, list = {next = 0xce09e8, prev = 0xce09e8}}
> > (gdb) n
> > libbpf: elf: skipping unrecognized data section(9) .eh_frame
> > libbpf: elf: skipping relo section(15) .rel.eh_frame for section(9) .eh_frame
> > 116 if (!skel) {
> > (gdb)
> > 121 skel->rodata->num_cpu = evsel__nr_cpus(evsel);
> > (gdb)
> > 123 bpf_map__resize(skel->maps.events, evsel__nr_cpus(evsel));
> > (gdb)
> > 124 bpf_map__resize(skel->maps.fentry_readings, 1);
> > (gdb)
> > 125 bpf_map__resize(skel->maps.accum_readings, 1);
> > (gdb)
> > 127 prog_name = bpf_target_prog_name(prog_fd);
> > (gdb)
> > 128 if (!prog_name) {
> > (gdb)
> > 133 bpf_object__for_each_program(prog, skel->obj) {
> > (gdb)
> > 134 err = bpf_program__set_attach_target(prog, prog_fd, prog_name);
> > (gdb)
> > 135 if (err) {
> > (gdb)
> > 133 bpf_object__for_each_program(prog, skel->obj) {
> > (gdb) p evsel
> > $12 = (struct evsel *) 0xce02c0
> > (gdb) p evsel->name
> > $13 = 0xce04e0 "cycles"
> > (gdb) n
> > 134 err = bpf_program__set_attach_target(prog, prog_fd, prog_name);
> > (gdb)
> > 135 if (err) {
> > (gdb)
> > 133 bpf_object__for_each_program(prog, skel->obj) {
> > (gdb)
> > 141 set_max_rlimit();
> > (gdb)
> > 142 err = bpf_prog_profiler_bpf__load(skel);
> > (gdb)
> > 143 if (err) {
> > (gdb)
> > 148 assert(skel != NULL);
> > (gdb)
> > 149 counter->skel = skel;
> > (gdb)
> > 150 list_add(&counter->list, &evsel->bpf_counter_list);
> > (gdb) c
> > Continuing.
> >
> > Breakpoint 4, bpf_program_profiler__install_pe (evsel=0xce02c0, cpu=0, fd=3) at util/bpf_counter.c:247
> > 247 {
> > (gdb) n
> > 252 list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
> > (gdb)
> > 253 skel = counter->skel;
> > (gdb) watch counter->skel
> > Hardware watchpoint 6: counter->skel
> > (gdb) p counter->skel
> > $14 = (void *) 0xce0a00
> > (gdb) n
> > 254 assert(skel != NULL);
> > (gdb) p skel
> > $15 = (struct bpf_prog_profiler_bpf *) 0xce0a00
> > (gdb) c
> > Continuing.
> >
> > Hardware watchpoint 6: counter->skel
> >
> > Old value = (void *) 0xce0a00
> > New value = (void *) 0x0
> > 0x00000000005cf45e in bpf_program_profiler__install_pe (evsel=0xce02c0, cpu=0, fd=3) at util/bpf_counter.c:252
> > 252 list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
>
> So it was the list operation that set counter->skel to NULL? I am really confused...

Yep, I'm confused as well and trying to reproduce this, but got
sidetracked...

- Arnaldo

> Song
>
> > (gdb) info b
> > Num Type Disp Enb Address What
> > 1 breakpoint keep y 0x00000000005ceb44 in bpf_program_profiler_load_one at util/bpf_counter.c:96
> > breakpoint already hit 1 time
> > 2 breakpoint keep y 0x00000000005cef2e in bpf_program_profiler__enable at util/bpf_counter.c:192
> > 3 breakpoint keep y 0x00000000005cf00a in bpf_program_profiler__read at util/bpf_counter.c:208
> > 4 breakpoint keep y 0x00000000005cf3ba in bpf_program_profiler__install_pe at util/bpf_counter.c:247
> > breakpoint already hit 1 time
> > 5 breakpoint keep y 0x00000000005ce8d4 in bpf_program_profiler__destroy at util/bpf_counter.c:47
> > 6 hw watchpoint keep y counter->skel
> > breakpoint already hit 1 time
> > (gdb) p evsel
> > $16 = (struct evsel *) 0xce02c0
> > (gdb) bt
> > #0 0x00000000005cf45e in bpf_program_profiler__install_pe (evsel=0xce02c0, cpu=0, fd=3) at util/bpf_counter.c:252
> > #1 0x00000000005cf4f3 in bpf_counter__install_pe (evsel=0xce02c0, cpu=0, fd=3) at util/bpf_counter.c:276
> > #2 0x00000000004f3ef1 in evsel__open_cpu (evsel=0xce02c0, cpus=0xce0500, threads=0xce75c0, start_cpu=0, end_cpu=1) at util/evsel.c:1792
> > #3 0x00000000004f4aa7 in evsel__open_per_cpu (evsel=0xce02c0, cpus=0xce0500, cpu=0) at util/evsel.c:1996
> > #4 0x000000000057f6f2 in create_perf_stat_counter (evsel=0xce02c0, config=0xa97860 <stat_config>, target=0xa97660 <target>, cpu=0) at util/stat.c:568
> > #5 0x0000000000435d96 in __run_perf_stat (argc=0, argv=0x7fffffffdb60, run_idx=0) at builtin-stat.c:790
> > #6 0x00000000004367e3 in run_perf_stat (argc=0, argv=0x7fffffffdb60, run_idx=0) at builtin-stat.c:1001
> > #7 0x0000000000439e35 in cmd_stat (argc=0, argv=0x7fffffffdb60) at builtin-stat.c:2415
> > #8 0x00000000004cf058 in run_builtin (p=0xaabe00 <commands+288>, argc=7, argv=0x7fffffffdb60) at perf.c:312
> > #9 0x00000000004cf2c5 in handle_internal_command (argc=7, argv=0x7fffffffdb60) at perf.c:364
> > #10 0x00000000004cf40c in run_argv (argcp=0x7fffffffd9ac, argv=0x7fffffffd9a0) at perf.c:408
> > #11 0x00000000004cf7d8 in main (argc=7, argv=0x7fffffffdb60) at perf.c:538
> > (gdb) n
> > 261 return 0;
> >
>

--

- Arnaldo