Do I need root (admin) permissions to run userspace 'perf' tool? (perf events are enabled in Linux kernel)

Question

Do I need to have to run perf userspace tool as system administrator (root), or can I run it (or at least some subcommands) as an ordinary user?

Solution

What you can do with perf without being root depends on the kernel.perf_event_paranoid sysctl setting.

• kernel.perf_event_paranoid = 2: you can't take any measurements. The perf utility might still be useful to analyse existing records with perf ls, perf report, perf timechart or perf trace.
• kernel.perf_event_paranoid = 1: you can trace a command with perf stat or perf record, and get kernel profiling data.
• kernel.perf_event_paranoid = 0: you can trace a command with perf stat or perf record, and get CPU event data.
• kernel.perf_event_paranoid = -1: you get raw access to kernel tracepoints (specifically, you can mmap the file created by perf_event_open, I don't know what the implications are).

Comments

[tcoolspy]

Unix programs that can't do what they need to do for lack of permissions will usually thrown an error if they can't do their job. Run it and see!

[Jakub Narębski]

I am asking this question to decide whether it is worth trying to install (as ordinary user, in $HOME) userspace part of perf tool (which is/can be non-trivial).

[Jakub Narębski]

FYI in Ubuntu perf is in linux-tools package, so installing perf there is simple.

[Martin Ueding]

@JakubNarębski: Except if it is not your own machine and the admins are reluctant to install packages.

[Jakub Narębski]

Nice. cat /proc/sys/kernel/perf_event_paranoid returns 1, so it seems that I would be able to take at least some measurements (BTW. whats the difference betwen "kernel profiling data" and "CPU event data"? Reference is enough)

[Gilles 'SO- stop being evil']

@Jakub: From what I understand, kernel events let you see calls to various kernel functions. CPU events are counters in the CPU that tell you how many times a particular location in memory was hit. I've never used them, so I can't tell you more about them; LWN has quite a few article on the topic, and it's still evolving.

[Peter Cordes]

With paranoid=2, you can still profile your own code in user-space (e.g. perf stat awk 'BEGIN{for(i=0;i<10000000;i++){}}' will show accurate user-space cycle and instruction counts, and you can even get counts for uops_issued.any and so on), but you don't get counts for code that ran during system calls/interrupts. So the reported CPU frequency (cycles / time) is at least slightly lower than actual because of time spent in kernel. See also What restriction is perf_event_paranoid == 1 actually putting on x86 perf?