| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Date: Tue, 19 Jan 2021 21:54:50 +0000
From: Song Liu <songliubraving@...com>
To: Arnaldo Carvalho de Melo <acme@...nel.org>
CC: open list <linux-kernel@...r.kernel.org>,
Peter Zijlstra <peterz@...radead.org>,
Ingo Molnar <mingo@...hat.com>,
Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
Namhyung Kim <namhyung@...nel.org>,
Mark Rutland <mark.rutland@....com>,
Jiri Olsa <jolsa@...hat.com>,
"Kernel Team" <Kernel-team@...com>
Subject: Re: [PATCH v7 3/3] perf-stat: enable counting events for BPF programs
> On Jan 19, 2021, at 8:31 AM, Arnaldo Carvalho de Melo <acme@...nel.org> 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@...nel.org> 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@...e perf]# perf trace -e 5sec.c
>> ^C
>> [root@...e 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@...e 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@...e ~]# 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...
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;
>
Powered by blists - more mailing lists