perf_events Frequently Asked Questions
The perf utility
Documentation and Development
Miscellaneous
Supported Machines
Q1a.
How can I download the perf utility?
A1a.
Initially to build perf you needed to download a full
linux-kernel source (or git) development tree. This is because
the perf utility code is not standalone, it includes
code from the actual linux kernel source.
This was not very convenient, so work was done to make it a bit
easier. Now that distributions are shipping kernels newer than
2.6.32 a lot of this pain has been taken away as pre-compiled
packages are available.
On Debian and Ubuntu the package to install is called
linux-tools-VER where VER is the kernel you are running.
On Fedora/RedHat the package is called perf
Note! on Debian/Ubuntu perf is actually a shell script
that calls an executable based on the current kernel version
( i.e. perf_2.6.32). If you upgrade the kernel by hand
you might have to run this executable directly or else get a nasty
error about using a mismatched perf/kernel. (In general any
perf executable should be backward compatible).
Q1b.
I have a specific perf question.
A1b.
I personally rarely use the perf utility except to verify
that perf_events is installed and working. Any questions
about the utility should be sent to the linux-perf-users
mailing list (see question 2a).
Q1c.
How do I measure Intel Offcore events with perf?
A1c.
As of the Linux 3.6 release the perf documentation
is very lacking in this area.
First be sure you have hardware that supports offcore
(generally Intel Nehalem or newer). You also need a recent kernel,
try Linux 3.6 or newer.
You need to find the
values for the events you are interested in. libpfm4 can help
with this. It will look something like:
perf stat -e cpu/cmask=1,event=2,umask=3,offcore_response=0x3f80408fff/ /bin/ls
Another way to do things is to find the values using libpfm4 check_events
program as such:
$ check_events OFFCORE_RESPONSE_0:ANY_DATA:REMOTE_DRAM
Supported PMU models:
.........
Codes : 0x5301b7 0x2033
And you can then use the results in perf like this:
perf stat -e cpu/config=0x5301b7,config1=0x2033,name=Remote_DRAM_Accesses/ ls
Q1d.
How do I measure Intel Uncore events with perf?
A1d.
As of the Linux 3.6 release the perf documentation
is very lacking in this area.
First be sure you have hardware that supports uncore
(generally Intel Nehalem or newer). AMD Northbridge events
are handled similarly but have various more complicated issues
(amd10h handled NB differently than newer chips, and there
was a kernel ABI break at one point).
You also need a recent kernel, try Linux 3.6 or newer.
For uncore support you'll want to be sure that in the
/sys/bus/event_source/devices/ directory there are some
uncore PMUs.
The syntax for doing an uncore call is something like this:
./perf stat -a -e "uncore_imc_0/event=0xff,umask=0x00/" /bin/ls
You may need root privileges to access uncore events.
Q2a.
Is there an official Perf Events mailing list?
A2a.
Typically all Perf Events development happens on the
main linux-kernel list with the appropriate maintainers cc'd.
This makes it very difficult for anyone trying to do perf_event
development who doesn't want to sift through a few thousand e-mails
a week.
A mailing list called
linux-perf-users exists but there is little traffic
and it's mainly about the perf userspace tool.
Q2b.
Is there good documentation for the perf events API?
A2b.
Unfortunately there isn't much official documentation, and
the documentation included with the perf tool is
currently out of date. The "best" documentation is reading
the kernel sources.
I have contributed a man page and some other
documentation here.
Q2c.
How do I tell if my kernel has perf_event support?
A2c.
The official answer is to look for the file
/proc/sys/kernel/perf_event_paranoid though that's only
there in 2.6.32 and newer (it is called perf_counter_paranoid in 2.6.31).
An old answer was to look for the
/sys/devices/system/cpu/perf_events
directory, but that was removed in the 2.6.37 kernel.
Q2d.
How do I get unsupported features in my new CPU into the
kernel?
A2d.
Create patches that implement your functionality.
Make sure these include support for the perf tool
and some attempt at creating "generalized" events that
use this feature, even if you don't plan on using them.
Your patch will likely not be considered otherwise.
(
See what happened when offcore event support
neglected this ).
Send your patches to the linux-kernel list, CCing the
maintainers listed for perf_events in the MAINTAINERS file.
Respond to any suggestions made by the developers.
If you're lucky your changes will be queued up for the next
merge window, meaning your changes will be in a released kernel
within the next 6 months to a year.
If you're unlucky your CPU will become obsolete before
support is added.
Q2e.
How do I determine the proper "raw" event value to
pass into perf when I want to use something other
than the limited set of perf "generalized" events?
A2e.
A good solution (short of reading the hardware manuals
and shifting the bits yourself) is to use
libpfm4.
The check_events and showevtinfo programs
available in the examples subdirectory can give you the
raw event code that you need.
Alternately, you can read the Intel Vol3b or AMD BKDG manuals
and convert the names into hex bitfields yourself, but that's
not much fun.
Q2f.
How do I use performance counters inside of the kernel?
A2f.
Be careful when doing this, as kernel usage of counters makes
them unavailable for users trying to do performance analysis.
You can look how the watchdog_nmi_enable() function does it
in the kernel/watchdog.c file in a recent linux kernel.
Q2g.
Can I set specific counters to measure
specific events with perf_event?
A2g.
perf_event was specifically designed to hide from you which
events end up in which counters.
You need to be careful when doing this, as the kernel could
be grabbing counters, and some counters have constraints on which
events can be measured in them.
In general, if you're not multiplexing, perf_event should be
somewhat consistent about which PMC gets which event.
If you really must do this, you can look into perfctr or perfmon2,
two different obsolete performance counter interfaces that
existed before perf_event. They allowed this level of control
over the counters. Their patches only apply to per-2.6.32 kernels though.
Q2i.
Can I use perf_events inside
of a virtualized environment
(such as KVM)?
A2i.
Yes, although the VM has to support it explicitly.
As of April 2012 you can get support on KVM if you use a
Linux 3.3 kernel on the host, and a current git-snapshot
of QEMU. Counter measurement works, overflow and profiling doesn't.
You can also measure the performance of the entire guest from
the outside using the "kvm" options to the perf tool.
Q2j.
How can I tell what low-level event the kernel
uses for a "generalized" event?
A2j.
On recent kernels kernels you can find this information using
the /sys filesystem. For example, if you want to find out
the parameters for the "instructions" event, you can
cat /sys/devices/cpu/events/instructions and
it will show you something like: event=0xc0
which you can look up in your chip's development manual
for more information.
For older kernels it was not so easy.
You needed to find the kernel source for the version of Linux you
are running (it has to be an exact match too, as the definitions
can and do change between kernel versions).
Then you need to dig around in the source code. For x86 chips,
the files you want are in arch/x86/kernel/cpu/.
For intel, look in perf_event_intel.c, for AMD look
in perf_event_amd.c, etc.
Then you have to find the definitions for the chip you are running.
For example, if it's a sandy bridge cache event, then look for
snb_hw_cache_event_ids. Then scroll down for the
event of interest. In this case, if you are interested in the
perf::DTLB-LOAD-MISSES event, it will be
with C(DTLB) / C(OP_READ) / C(RESULT_MISS), which maps
to 0x0108. If you're lucky there will be a comment telling
you want event this maps to (though you might have to dig
out the architectural manuals to make sure there's not a typo,
which also happens occasionally). In this case it lists
DTLB_LOAD_MISSES.CAUSES_A_WALK.
Similar methodology can be used to find events on other architectures,
although you might need to use grep to find out exactly which file
has the definitions. You also might need to play around with
the perf_event.h file to figure out which "generalized" names match
to which #defines in the event arrays. And finally, you should
verify using perf that the results you get using
the event you think you figured out matches the one when using
the generalized event.
Q3a.
Do you recommend using the PAPI perf_event substrate?
A3a.
The combination of PAPI 4.2 (or newer) and
Linux 2.6.34 (or newer) can give you a
reasonable PAPI experience.
Perfmon2 development stopped before Linux 2.6.30 and perfctr
development has slowed, so perf_event might be your only option
going forward.
Do note that the linux-kernel perf_event development is slow,
so new processor support and new counting methodologies
(such as uncore or LWP) can take years before they show
up in a vendor kernel.
Q3b.
Are you some sort of
perfmon2
contributor troll
who hates perf_events?
A3b.
I think by this point I have had more code committed to
the perf events tree than I ever did to the perfmon2 tree.
My main disagreement with perf events (besides the re-inventing
of the wheel involved that's cost me a lot of development time)
is the current policy of putting event name mappings in the kernel.
I believe this belongs in a user-space library that can be
easily updated without a kernel upgrade. The main perf events
developers strongly disagree.
Q4a.
Does my CPU support perf_events?
A4a.
See the CPU compatibility matrix here.
Q4b.
Are counters supported on the Raspberry Pi?
A4b.
There should be support by 3.16 or 3.17.
The Raspberry Pi does have counter hardware, but
the PMU interrupt is not connected so perf_event did not enable
the interface.
For details of the support see our document
here
Back to the unofficial perf_events page