[PATCH v8 00/12] Introduce CAP_PERFMON to secure system performance monitoring and observability

Arnaldo Carvalho de Melo arnaldo.melo at gmail.com
Tue Apr 7 16:56:43 UTC 2020


Em Tue, Apr 07, 2020 at 01:36:54PM -0300, Arnaldo Carvalho de Melo escreveu:
> Em Tue, Apr 07, 2020 at 05:54:27PM +0300, Alexey Budankov escreveu:
> > On 07.04.2020 17:35, Arnaldo Carvalho de Melo wrote:
> > > Em Tue, Apr 07, 2020 at 11:30:14AM -0300, Arnaldo Carvalho de Melo escreveu:
> > >> [perf at five ~]$ type perf
> > >> perf is hashed (/home/perf/bin/perf)
> > >> [perf at five ~]$ getcap /home/perf/bin/perf
> > >> /home/perf/bin/perf = cap_sys_ptrace,cap_syslog,38+ep
> > >> [perf at five ~]$ groups
> > >> perf perf_users
> > >> [perf at five ~]$ id
> > >> uid=1002(perf) gid=1002(perf) groups=1002(perf),1003(perf_users) context=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023
> > >> [perf at five ~]$ perf top --stdio
> > >> Error:
> > >> Failed to mmap with 1 (Operation not permitted)
> > >> [perf at five ~]$ perf record -a
> > >> ^C[ perf record: Woken up 1 times to write data ]
> > >> [ perf record: Captured and wrote 1.177 MB perf.data (1552 samples) ]
> > >>
> > >> [perf at five ~]$ perf evlist
> > >> cycles:u
> > >> [perf at five ~]$
> > > 
> > > Humm, perf record falls back to cycles:u after initially trying cycles
> > > (i.e. kernel and userspace), lemme see trying 'perf top -e cycles:u',
> > > lemme test, humm not really:
> > > 
> > > [perf at five ~]$ perf top --stdio -e cycles:u
> > > Error:
> > > Failed to mmap with 1 (Operation not permitted)
> > > [perf at five ~]$ perf record -e cycles:u -a sleep 1
> > > [ perf record: Woken up 1 times to write data ]
> > > [ perf record: Captured and wrote 1.123 MB perf.data (132 samples) ]
> > > [perf at five ~]$
> > > 
> > > Back to debugging this.
> > 
> > Could makes sense adding cap_ipc_lock to the binary to isolate from this:
> > 
> > kernel/events/core.c: 6101
> > 	if ((locked > lock_limit) && perf_is_paranoid() &&
> > 		!capable(CAP_IPC_LOCK)) {
> > 		ret = -EPERM;
> > 		goto unlock;
> > 	}
> 
> 
> That did the trick, I'll update the documentation and include in my
> "Committer testing" section:
> 
> [perf at five ~]$ groups
> perf perf_users
> [perf at five ~]$ ls -lahF bin/perf
> -rwxr-x---. 1 root perf_users 24M Apr  7 10:34 bin/perf*
> [perf at five ~]$ getcap bin/perf
> bin/perf = cap_ipc_lock,cap_sys_ptrace,cap_syslog,38+ep
> [perf at five ~]$
> [perf at five ~]$ perf top --stdio
> 
> 
>    PerfTop:     652 irqs/sec  kernel:73.8%  exact: 99.7% lost: 0/0 drop: 0/0 [4000Hz cycles:u],  (all, 12 CPUs)
> ---------------------------------------------------------------------------------------------------------------
> 
>     13.03%  [kernel]               [k] module_get_kallsym
>      5.25%  [kernel]               [k] kallsyms_expand_symbol.constprop.0
>      5.00%  libc-2.30.so           [.] __GI_____strtoull_l_internal
>      4.41%  [kernel]               [k] memcpy
>      3.42%  [kernel]               [k] vsnprintf
>      2.98%  perf                   [.] map__process_kallsym_symbol
>      2.86%  [kernel]               [k] format_decode
>      2.73%  [kernel]               [k] number
>      2.70%  perf                   [.] rb_next
>      2.59%  perf                   [.] maps__split_kallsyms
>      2.54%  [kernel]               [k] string_nocheck
>      1.90%  libc-2.30.so           [.] _IO_getdelim
>      1.86%  [kernel]               [k] __x86_indirect_thunk_rax
>      1.53%  libc-2.30.so           [.] _int_malloc
>      1.48%  libc-2.30.so           [.] __memmove_avx_unaligned_erms
>      1.40%  [kernel]               [k] clear_page_rep
>      1.07%  perf                   [.] rb_insert_color
>      1.01%  libc-2.30.so           [.] _IO_feof
>      0.99%  perf                   [.] __dso__load_kallsyms
>      0.98%  [kernel]               [k] s_next
>      0.96%  perf                   [.] __rblist__findnew
>      0.95%  [kernel]               [k] strlen
>      0.95%  perf                   [.] arch__symbols__fixup_end
>      0.94%  libpixman-1.so.0.38.4  [.] 0x000000000006f4af
>      0.94%  perf                   [.] symbol__new
>      0.89%  libpixman-1.so.0.38.4  [.] 0x000000000006f4a0
>      0.86%  [kernel]               [k] seq_read
>      0.81%  libpixman-1.so.0.38.4  [.] 0x000000000006f4ab
>      0.80%  perf                   [.] __symbols__insert
>      0.73%  libpixman-1.so.0.38.4  [.] 0x000000000006f4a7
>      0.67%  [kernel]               [k] s_show
>      0.66%  libc-2.30.so           [.] __libc_calloc
>      0.61%  libpixman-1.so.0.38.4  [.] 0x000000000006f4bb
>      0.59%  [kernel]               [k] get_page_from_freelist
>      0.59%  perf                   [.] memcpy at plt
>      0.58%  perf                   [.] eprintf
> exiting.
> [perf at five ~]$
> 
> There is still something strange in here, the event is cycles:u (see at
> the PerfTop line, but it is getting kernel samples :-\

So running with 'perf top --stdio -vv 2> /tmp/output' I see we try
create three events, the first is some capability querying, then we try
to determine the max precision level, but continue with
attr.exclude_kernel=1, which shouldn't be the case, perhaps we're seeing
that it is not the root in the tooling part, and end up setting that to
1 as, previously, we knew it would fail, so we should switch to checking
if we have cap_perfmon too, will check that:

------------------------------------------------------------
perf_event_attr:
  type                             1
  size                             120
  config                           0x9
  watermark                        1
  sample_id_all                    1
  bpf_event                        1
  { wakeup_events, wakeup_watermark } 1
------------------------------------------------------------
------------------------------------------------------------
perf_event_attr:
  size                             120
  { sample_period, sample_freq }   4000
  sample_type                      IP|TID|TIME|CPU|PERIOD
  read_format                      ID
  disabled                         1
  inherit                          1
  exclude_kernel                   1
  mmap                             1
  comm                             1
  freq                             1
  task                             1
  precise_ip                       3
  sample_id_all                    1
  exclude_guest                    1
  mmap2                            1
  comm_exec                        1
  ksymbol                          1
  bpf_event                        1
------------------------------------------------------------
------------------------------------------------------------
perf_event_attr:
  size                             120
  { sample_period, sample_freq }   4000
  sample_type                      IP|TID|TIME|CPU|PERIOD
  read_format                      ID
  disabled                         1
  inherit                          1
  exclude_kernel                   1
  mmap                             1
  comm                             1
  freq                             1
  task                             1
  precise_ip                       2
  sample_id_all                    1
  exclude_guest                    1
  mmap2                            1
  comm_exec                        1
  ksymbol                          1
  bpf_event                        1
------------------------------------------------------------

But then, even with that attr.exclude_kernel set to 1 we _still_ get
kernel samples, which looks like another bug, now trying with strace,
which leads us to another rabbit hole:

[perf at five ~]$ strace -e perf_event_open -o /tmp/out.put perf top --stdio
Error:
You may not have permission to collect system-wide stats.

Consider tweaking /proc/sys/kernel/perf_event_paranoid,
which controls use of the performance events system by
unprivileged users (without CAP_PERFMON or CAP_SYS_ADMIN).

The current value is 2:

  -1: Allow use of (almost) all events by all users
      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK
>= 0: Disallow ftrace function tracepoint by users without CAP_PERFMON or CAP_SYS_ADMIN
      Disallow raw tracepoint access by users without CAP_SYS_PERFMON or CAP_SYS_ADMIN
>= 1: Disallow CPU event access by users without CAP_PERFMON or CAP_SYS_ADMIN
>= 2: Disallow kernel profiling by users without CAP_PERFMON or CAP_SYS_ADMIN

To make this setting permanent, edit /etc/sysctl.conf too, e.g.:

	kernel.perf_event_paranoid = -1

[perf at five ~]$

If I remove that strace -e ... from the front, 'perf top' is back
working as a non-cap_sys_admin user, just with cap_perfmon.

- Arnaldo



More information about the Linux-security-module-archive mailing list