RE: [PATCH v3 3/4] perf: Use CAP_SYSLOG with kptr_restrict checks

["Lubashev, Igor" <ilubashe@akamai.com>]

On Mon, August 19, 2019 at 12:51 PM Mathieu Poirier <mathieu.poirier@linaro.org> wrote:
> On Thu, 15 Aug 2019 at 15:42, Arnaldo Carvalho de Melo
> <arnaldo.melo@gmail.com> wrote:
> >
> > Em Thu, Aug 15, 2019 at 02:16:48PM -0600, Mathieu Poirier escreveu:
> > > On Wed, 14 Aug 2019 at 14:02, Lubashev, Igor <ilubashe@akamai.com>
> wrote:
> > > >
> > > > > On Wed, August 14, 2019 at 2:52 PM Arnaldo Carvalho de Melo
> <arnaldo.melo@gmail.com> wrote:
> > > > > Em Wed, Aug 14, 2019 at 03:48:14PM -0300, Arnaldo Carvalho de
> > > > > Melo
> > > > > escreveu:
> > > > > > Em Wed, Aug 14, 2019 at 12:04:33PM -0600, Mathieu Poirier
> escreveu:
> > > > > > > # echo 0 > /proc/sys/kernel/kptr_restrict #
> > > > > > > ./tools/perf/perf record -e instructions:k uname
> > > > > > > perf: Segmentation fault
> > > > > > > Obtained 10 stack frames.
> > > > > > > ./tools/perf/perf(sighandler_dump_stack+0x44)
> > > > > > > [0x55af9e5da5d4]
> > > > > > > /lib/x86_64-linux-gnu/libc.so.6(+0x3ef20) [0x7fd31efb6f20]
> > > > > > > ./tools/perf/perf(perf_event__synthesize_kernel_mmap+0xa7)
> > > > > > > [0x55af9e590337]
> > > > > > > ./tools/perf/perf(+0x1cf5be) [0x55af9e50c5be]
> > > > > > > ./tools/perf/perf(cmd_record+0x1022) [0x55af9e50dff2]
> > > > > > > ./tools/perf/perf(+0x23f98d) [0x55af9e57c98d]
> > > > > > > ./tools/perf/perf(+0x23fc9e) [0x55af9e57cc9e]
> > > > > > > ./tools/perf/perf(main+0x369) [0x55af9e4f6bc9]
> > > > > > > /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xe7)
> > > > > > > [0x7fd31ef99b97]
> > > > > > > ./tools/perf/perf(_start+0x2a) [0x55af9e4f704a] Segmentation
> > > > > > > fault
> > > > > > >
> > > > > > > I can reproduce this on both x86 and ARM64.
> > > > > >
> > > > > > I don't see this with these two csets removed:
> > > > > >
> > > > > > 7ff5b5911144 perf symbols: Use CAP_SYSLOG with kptr_restrict
> > > > > > checks d7604b66102e perf tools: Use CAP_SYS_ADMIN with
> > > > > > perf_event_paranoid checks
> > > > > >
> > > > > > Which were the ones I guessed were related to the problem you
> > > > > > reported, so they are out of my ongoing perf/core pull request
> > > > > > to Ingo/Thomas, now trying with these applied and your
> instructions...
> > > > >
> > > > > Can't repro:
> > > > >
> > > > > [root@quaco ~]# cat /proc/sys/kernel/kptr_restrict
> > > > > 0
> > > > > [root@quaco ~]# perf record -e instructions:k uname Linux [ perf
> record:
> > > > > Woken up 1 times to write data ] [ perf record: Captured and
> > > > > wrote 0.024 MB perf.data (1 samples) ] [root@quaco ~]# echo 1 >
> > > > > /proc/sys/kernel/kptr_restrict [root@quaco ~]# perf record -e
> > > > > instructions:k uname Linux [ perf record: Woken up 1 times to write
> data ] [ perf record:
> > > > > Captured and wrote 0.024 MB perf.data (1 samples) ] [root@quaco
> > > > > ~]# echo
> > > > > 0 > /proc/sys/kernel/kptr_restrict [root@quaco ~]# perf record
> > > > > -e instructions:k uname Linux [ perf record: Woken up 1 times to
> > > > > write data ] [ perf record: Captured and wrote 0.024 MB
> > > > > perf.data (1 samples) ] [root@quaco ~]#
> > > > >
> > > > > [acme@quaco perf]$ git log --oneline --author Lubashev tools/
> > > > > 7ff5b5911144 (HEAD -> perf/cap, acme.korg/tmp.perf/cap,
> > > > > acme.korg/perf/cap) perf symbols: Use CAP_SYSLOG with
> > > > > kptr_restrict checks d7604b66102e perf tools: Use CAP_SYS_ADMIN
> > > > > with perf_event_paranoid checks c766f3df635d perf ftrace: Use
> > > > > CAP_SYS_ADMIN instead of euid==0 c22e150e3afa perf tools: Add
> > > > > helpers to use capabilities if present
> > > > > 74d5f3d06f70 tools build: Add capability-related feature
> > > > > detection perf version 5.3.rc4.g7ff5b5911144 [acme@quaco perf]$
> > > >
> > > > I got an ARM64 cloud VM, but I cannot reproduce.
> > > > # cat /proc/sys/kernel/kptr_restrict
> > > > 0
> > > >
> > > > Perf trace works fine (does not die):
> > > > # ./perf trace -a
> > > >
> > > > Here is my setup:
> > > > Repo: git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git
> > > > Branch: tmp.perf/cap
> > > > Commit: 7ff5b5911 "perf symbols: Use CAP_SYSLOG with kptr_restrict
> checks"
> > > > gcc --version: gcc (Ubuntu/Linaro 7.4.0-1ubuntu1~18.04.1) 7.4.0
> > > > uname -a: Linux arm-4-par-1 4.9.93-mainline-rev1 #1 SMP Tue Apr 10
> > > > 09:54:46 UTC 2018 aarch64 aarch64 aarch64 GNU/Linux lsb_release
> > > > -a: Ubuntu 18.04.3 LTS
> > > >
> > > > Auto-detecting system features:
> > > > ...                         dwarf: [ on  ]
> > > > ...            dwarf_getlocations: [ on  ]
> > > > ...                         glibc: [ on  ]
> > > > ...                          gtk2: [ on  ]
> > > > ...                      libaudit: [ on  ]
> > > > ...                        libbfd: [ on  ]
> > > > ...                        libcap: [ on  ]
> > > > ...                        libelf: [ on  ]
> > > > ...                       libnuma: [ on  ]
> > > > ...        numa_num_possible_cpus: [ on  ]
> > > > ...                       libperl: [ on  ]
> > > > ...                     libpython: [ on  ]
> > > > ...                     libcrypto: [ on  ]
> > > > ...                     libunwind: [ on  ]
> > > > ...            libdw-dwarf-unwind: [ on  ]
> > > > ...                          zlib: [ on  ]
> > > > ...                          lzma: [ on  ]
> > > > ...                     get_cpuid: [ OFF ]
> > > > ...                           bpf: [ on  ]
> > > > ...                        libaio: [ on  ]
> > > > ...                       libzstd: [ on  ]
> > > > ...        disassembler-four-args: [ on  ]
> > > >
> > > > I also could not reproduce on x86:
> > > > lsb_release -a: Ubuntu 18.04.1 LTS gcc --version: gcc (Ubuntu
> > > > 7.4.0-1ubuntu1~18.04aka10.0.0) 7.4.0 uname -r: 4.4.0-154-generic
> > >
> > > I isolated the problem to libcap-dev - if it is not installed a
> > > segmentation fault will occur.  Since this set is about using
> > > capabilities it is obvious that not having it on a system should
> > > fail a trace session, but it should not crash it.
> >
> > It shouldn't crash on x86_64:
> >
> > root@quaco ~]# sysctl kernel.kptr_restrict kernel.kptr_restrict = 0
> > [root@quaco ~]# perf record -e instructions:k uname Linux [ perf
> > record: Woken up 1 times to write data ] [ perf record: Captured and
> > wrote 0.023 MB perf.data (5 samples) ] [root@quaco ~]# ldd ~/bin/perf
> > | grep libcap [root@quaco ~]# perf -v perf version
> > 5.3.rc4.g329ca330bf8b [root@quaco ~]# [acme@quaco perf]$ git log
> > --oneline -4 329ca330bf8b (HEAD -> perf/cap) perf symbols: Use
> > CAP_SYSLOG with kptr_restrict checks f7b9999387af perf tools: Use
> > CAP_SYS_ADMIN with perf_event_paranoid checks
> > 4d0489cf1c47 (acme.korg/tmp.perf/script-switch-on-off, perf/core) perf
> > report: Add --switch-on/--switch-off events
> > 2f53ae347f59 perf top: Add --switch-on/--switch-off events [acme@quaco
> > perf]$
> >
> > > If libcap-dev is not installed function
> > > symbol__restricted_filename() will return true, which in turn will
> > > prevent symbol_name to be set in
> > > machine__get_running_kernel_start().  That prevents function
> > > map__set_kallsyms_ref_reloc_sym() from being called in
> > > machine__create_kernel_maps(), resulting in kmap->ref_reloc_sym
> > > being NULL in _perf_event__synthesize_kernel_mmap() and a
> > > segmentation fault.
> >
> > Can you please try with my perf/cap branch? Perhaps you're using
> > Igor's original set of patches? I made changes to the fallback, he was
> > making it return false while I made it fallback to geteuid() == 0, as
> > was before his patches.
> 
> Things are working properly on your perf/cap branch.  I tested with on both
> x86 and ARM.

Mathieu, you are probably testing with euid==0.  If you were to test with euid!=0 but with CAP_SYSLOG and no libcap (and kptr_restrict=0, perf_event_paranoid=2), you would likely hit the bug that you identified in __perf_event__synthesize_kermel_mmap().

See https://lkml.kernel.org/lkml/930a59730c0d495f8c5acf4f99048e8d@usma1ex-dag1mb6.msg.corp.akamai.com for the fix (Option 1 only or Options 1+2).

Arnaldo, once we decide what the right fix is, I am happy to post the update (options 1, 1+2) as a patch series.

- Igor


> > > I am not sure how this can be fixed.  I counted a total of 19
> > > instances where kmap->ref_reloc_sym->XYZ is called, only 2 of wich
> > > care to check if kmap->ref_reloc_sym is valid before proceeding.  As
> > > such I must hope that in the 17 other cases, kmap->ref_reloc_sym is
> > > guaranteed to be valid.  If I am correct then all we need is to
> > > check for a valid pointer in _perf_event__synthesize_kernel_mmap().
> > > Otherwise it will be a little harder.
> > >
> > > Mathieu