Hi,
All the other perf tooling can be used as a normal user. At one point,
I had to touch /proc/sys/kernel/perf_event_paranoid to get me numbers
on a PMU event without a symbolic name, but that's about it.
perf-trace seems to be the exception though:
artagnon|perf-core=:~/src/linux/tools/perf$ perf trace -a -e open
Couldn't read the raw_syscalls tracepoints information!
The responsible codepath is:
if (perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_enter",
trace__sys_enter) ||
perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_exit",
trace__sys_exit)) {
I don't understand much of what's going on, but what can be done to
improve the situation for users?
Thanks.
On 10/2/13 3:20 AM, Ramkumar Ramachandra wrote:
> Hi,
>
> All the other perf tooling can be used as a normal user. At one point,
> I had to touch /proc/sys/kernel/perf_event_paranoid to get me numbers
> on a PMU event without a symbolic name, but that's about it.
> perf-trace seems to be the exception though:
>
> artagnon|perf-core=:~/src/linux/tools/perf$ perf trace -a -e open
> Couldn't read the raw_syscalls tracepoints information!
>
> The responsible codepath is:
>
> if (perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_enter",
> trace__sys_enter) ||
> perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_exit",
> trace__sys_exit)) {
>
> I don't understand much of what's going on, but what can be done to
> improve the situation for users?
If you trust your users make the debugfs mount point rx by group,world.
David
David Ahern wrote:
> If you trust your users make the debugfs mount point rx by group,world.
Thanks David. I can preserve this configuration across reboots by
putting an entry in fstab, right? How do I preserve the value of
/proc/sys/kernel/perf_event_paranoid across reboots?
* Ramkumar Ramachandra <[email protected]> wrote:
> David Ahern wrote:
> > If you trust your users make the debugfs mount point rx by group,world.
>
> Thanks David. I can preserve this configuration across reboots by
> putting an entry in fstab, right? How do I preserve the value of
> /proc/sys/kernel/perf_event_paranoid across reboots?
via:
echo 'kernel.perf_event_paranoid = -1' >> /etc/sysctl.conf
I think 'perf trace' should probably print such suggestions when it
notices a privilege problem, to make it far more obvious for new users to
correctly configure their system for easy tracing.
The current output:
comet:~/tip> perf trace
Couldn't read the raw_syscalls tracepoints information!
is as hostile to the user as it gets ;-)
Thanks,
Ingo
Ingo Molnar wrote:
> echo 'kernel.perf_event_paranoid = -1' >> /etc/sysctl.conf
Thanks Ingo. It's called /etc/sysctl.d/* now:
https://www.archlinux.org/news/deprecation-of-etcsysctlconf/
> The current output:
>
> comet:~/tip> perf trace
> Couldn't read the raw_syscalls tracepoints information!
>
> is as hostile to the user as it gets ;-)
Patches are on the way :)
* Ramkumar Ramachandra <[email protected]> wrote:
> Ingo Molnar wrote:
> > echo 'kernel.perf_event_paranoid = -1' >> /etc/sysctl.conf
>
> Thanks Ingo. It's called /etc/sysctl.d/* now:
> https://www.archlinux.org/news/deprecation-of-etcsysctlconf/
That works too, on modern distros.
Thanks,
Ingo
Em Thu, Oct 03, 2013 at 08:25:05AM +0200, Ingo Molnar escreveu:
> * Ramkumar Ramachandra <[email protected]> wrote:
> > David Ahern wrote:
> > > If you trust your users make the debugfs mount point rx by group,world.
> > Thanks David. I can preserve this configuration across reboots by
> > putting an entry in fstab, right? How do I preserve the value of
> > /proc/sys/kernel/perf_event_paranoid across reboots?
> via:
> echo 'kernel.perf_event_paranoid = -1' >> /etc/sysctl.conf
You don't need that, only if you want users to trace syswide or threads
from other users:
[acme@zoo linux]$ cat /proc/sys/kernel/perf_event_paranoid
1
[acme@zoo linux]$ trace -e brk usleep 1
0.420 ( 0.001 ms): brk( ) = 0x2427000
0.698 ( 0.002 ms): brk( ) = 0x2427000
0.702 ( 0.002 ms): brk(brk: 0x2448000 ) = 0x2448000
0.704 ( 0.001 ms): brk( ) = 0x2448000
[acme@zoo linux]$
When doing syswide we get another message that can get some love and care:
[acme@zoo linux]$ trace -a
Couldn't create the events: Operation not permitted
Here we should tell that if this is done, as you suggested above:
[root@zoo ~]# echo -1 > /proc/sys/kernel/perf_event_paranoid
Then lets do syswide tracing filtering out common syscalls:
[acme@zoo linux]$ trace -a -e \!write,lseek,ioctl,recvfrom,read,open,getrlimit,writev,poll,rt_sigprocmask,brk,,timer_gettime,timer_settime,gettid,setitimer,select,rt_sigaction,nanosleep,newfstatat,stat,close,openat,getdents,recvmsg,epoll_wait,sendmsg,futex,mmap,munmap,getsockopt,getegid,geteuid,lstat,fstat,inotify_add_watch,exit,madvise,fcntl,alarm,exit_group,wait4,unlink,link,dup,sendto,getuid,setgid,uname,umask,getgid,fsync,chdir,setuid,setreuid,setregid,socket,connect,access,mprotect,getpgrp,setresgid,getpid,execve,dup2,clone,getsockname,bind,arch_prctl,statfs,getppid,pipe2,capget,readlink,fstatfs
90.964 ( 0.001 ms): python/7597 set_robust_list(head: 0x7ff5cf2d09e0, len: 24 ) = 0
1092.945 ( 0.002 ms): python/7598 set_robust_list(head: 0x7ff5cf2d09e0, len: 24 ) = 0
1104.372 ( 0.004 ms): fetchmail/2333 setresuid(ruid: -1, euid: 1000, suid: -1 ) = 0
1104.589 ( 0.003 ms): fetchmail/2333 setresuid(ruid: -1, euid: 1000, suid: -1 ) = 0
1104.869 ( 0.002 ms): fetchmail/7599 set_robust_list(head: 0x7f4d97395ae0, len: 24 ) = 0
2095.270 ( 0.001 ms): python/7600 set_robust_list(head: 0x7ff5cf2d09e0, len: 24 ) = 0
2581.513 ( 0.004 ms): fetchmail/2333 setresuid(ruid: -1, euid: 1000, suid: -1 ) = 0
2581.676 ( 0.002 ms): fetchmail/7601 set_robust_list(head: 0x7f4d97395ae0, len: 24 ) = 0
2581.688 ( 0.003 ms): fetchmail/2333 setresuid(ruid: -1, euid: 1000, suid: -1 ) = 0
3097.509 ( 0.002 ms): python/7602 set_robust_list(head: 0x7ff5cf2d09e0, len: 24 ) = 0
4100.418 ( 0.001 ms): python/7603 set_robust_list(head: 0x7ff5cf2d09e0, len: 24 ) = 0
^C[acme@zoo linux]$
> I think 'perf trace' should probably print such suggestions when it
> notices a privilege problem, to make it far more obvious for new users to
> correctly configure their system for easy tracing.
>
> The current output:
>
> comet:~/tip> perf trace
> Couldn't read the raw_syscalls tracepoints information!
>
> is as hostile to the user as it gets ;-)
Yes, its good that there are users now! ;-)
- Arnaldo