Measuring large profoundly-idle configurations
requires turbostat to be more lightweight.
Otherwise, the operation of turbostat itself
can interfere with the measurements.
This re-write makes turbostat topology aware.
Hardware is accessed in "topology order".
Redundant hardware accesses are deleted.
Redundant output is deleted.
Also, output is buffered and
local RDTSC use replaces remote MSR access for TSC.
From a feature point of view, the output
looks different since redundant figures are absent.
Also, there are now -c and -p options -- to restrict
output to the 1st thread in each core, and the 1st
thread in each package, respectively. This is helpful
to reduce output on big systems, where more detail
than the "-s" system summary is desired.
Finally, periodic mode output is now on stdout, not stderr.
Turbostat v2 is also slightly more robust in
handling run-time CPU online/offline events,
as it now checks the actual map of on-line cpus rather
than just the total number of on-line cpus.
Signed-off-by: Len Brown <len.brown@intel.com>
Initial IVB support went into turbostat in Linux-3.1:
553575f1ae
(tools turbostat: recognize and run properly on IVB)
However, when running on IVB, turbostat would fail
to report the new couters added with SNB, c7, pc2 and pc7.
So in scenarios where these counters are non-zero on IVB,
turbostat would report erroneous residencey results.
In particular c7 time would be added to c1 time,
since c1 time is calculated as "that which is left over".
Also, turbostat reports MHz capabilities when passed
the "-v" option, and it would incorrectly report 133MHz
bclk instead of 100MHz bclk for IVB, which would inflate
GHz reported with that option.
This patch is a backport of a fix already included in turbostat v2.
Signed-off-by: Len Brown <len.brown@intel.com>
Linux 3.4 included a modification to turbostat to
lower cross-call overhead by using scheduler affinity:
15aaa34654
(tools turbostat: reduce measurement overhead due to IPIs)
In the use-case where turbostat forks a child program,
that change had the un-intended side-effect of binding
the child to the last cpu in the system.
This change removed the binding before forking the child.
This is a back-port of a fix already included in turbostat v2.
Signed-off-by: Len Brown <len.brown@intel.com>
It's been broken forever (i.e. it's not scheduling in a power
aware fashion), as reported by Suresh and others sending
patches, and nobody cares enough to fix it properly ...
so remove it to make space free for something better.
There's various problems with the code as it stands today, first
and foremost the user interface which is bound to topology
levels and has multiple values per level. This results in a
state explosion which the administrator or distro needs to
master and almost nobody does.
Furthermore large configuration state spaces aren't good, it
means the thing doesn't just work right because it's either
under so many impossibe to meet constraints, or even if
there's an achievable state workloads have to be aware of
it precisely and can never meet it for dynamic workloads.
So pushing this kind of decision to user-space was a bad idea
even with a single knob - it's exponentially worse with knobs
on every node of the topology.
There is a proposal to replace the user interface with a single
3 state knob:
sched_balance_policy := { performance, power, auto }
where 'auto' would be the preferred default which looks at things
like Battery/AC mode and possible cpufreq state or whatever the hw
exposes to show us power use expectations - but there's been no
progress on it in the past many months.
Aside from that, the actual implementation of the various knobs
is known to be broken. There have been sporadic attempts at
fixing things but these always stop short of reaching a mergable
state.
Therefore this wholesale removal with the hopes of spurring
people who care to come forward once again and work on a
coherent replacement.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/1326104915.2442.53.camel@twins
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull ACPI & Power Management changes from Len Brown:
- ACPI 5.0 after-ripples, ACPICA/Linux divergence cleanup
- cpuidle evolving, more ARM use
- thermal sub-system evolving, ditto
- assorted other PM bits
Fix up conflicts in various cpuidle implementations due to ARM cpuidle
cleanups (ARM at91 self-refresh and cpu idle code rewritten into
"standby" in asm conflicting with the consolidation of cpuidle time
keeping), trivial SH include file context conflict and RCU tracing fixes
in generic code.
* 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/lenb/linux: (77 commits)
ACPI throttling: fix endian bug in acpi_read_throttling_status()
Disable MCP limit exceeded messages from Intel IPS driver
ACPI video: Don't start video device until its associated input device has been allocated
ACPI video: Harden video bus adding.
ACPI: Add support for exposing BGRT data
ACPI: export acpi_kobj
ACPI: Fix logic for removing mappings in 'acpi_unmap'
CPER failed to handle generic error records with multiple sections
ACPI: Clean redundant codes in scan.c
ACPI: Fix unprotected smp_processor_id() in acpi_processor_cst_has_changed()
ACPI: consistently use should_use_kmap()
PNPACPI: Fix device ref leaking in acpi_pnp_match
ACPI: Fix use-after-free in acpi_map_lsapic
ACPI: processor_driver: add missing kfree
ACPI, APEI: Fix incorrect APEI register bit width check and usage
Update documentation for parameter *notrigger* in einj.txt
ACPI, APEI, EINJ, new parameter to control trigger action
ACPI, APEI, EINJ, limit the range of einj_param
ACPI, APEI, Fix ERST header length check
cpuidle: power_usage should be declared signed integer
...
Sometimes users have turbostat running in interval mode
when they take processors offline/online.
Previously, turbostat would survive, but not gracefully.
Tighten up the error checking so turbostat notices
changesn sooner, and print just 1 line on change:
turbostat: re-initialized with num_cpus %d
Signed-off-by: Len Brown <len.brown@intel.com>
turbostat uses /dev/cpu/*/msr interface to read MSRs.
For modern systems, it reads 10 MSR/CPU. This can
be observed as 10 "Function Call Interrupts"
per CPU per sample added to /proc/interrupts.
This overhead is measurable on large idle systems,
and as Yoquan Song pointed out, it can even trick
cpuidle into thinking the system is busy.
Here turbostat re-schedules itself in-turn to each
CPU so that its MSR reads will always be local.
This replaces the 10 "Function Call Interrupts"
with a single "Rescheduling interrupt" per sample
per CPU.
On an idle 32-CPU system, this shifts some residency from
the shallow c1 state to the deeper c7 state:
# ./turbostat.old -s
%c0 GHz TSC %c1 %c3 %c6 %c7 %pc2 %pc3 %pc6 %pc7
0.27 1.29 2.29 0.95 0.02 0.00 98.77 20.23 0.00 77.41 0.00
0.25 1.24 2.29 0.98 0.02 0.00 98.75 20.34 0.03 77.74 0.00
0.27 1.22 2.29 0.54 0.00 0.00 99.18 20.64 0.00 77.70 0.00
0.26 1.22 2.29 1.22 0.00 0.00 98.52 20.22 0.00 77.74 0.00
0.26 1.38 2.29 0.78 0.02 0.00 98.95 20.51 0.05 77.56 0.00
^C
i# ./turbostat.new -s
%c0 GHz TSC %c1 %c3 %c6 %c7 %pc2 %pc3 %pc6 %pc7
0.27 1.20 2.29 0.24 0.01 0.00 99.49 20.58 0.00 78.20 0.00
0.27 1.22 2.29 0.25 0.00 0.00 99.48 20.79 0.00 77.85 0.00
0.27 1.20 2.29 0.25 0.02 0.00 99.46 20.71 0.03 77.89 0.00
0.28 1.26 2.29 0.25 0.01 0.00 99.46 20.89 0.02 77.67 0.00
0.27 1.20 2.29 0.24 0.01 0.00 99.48 20.65 0.00 78.04 0.00
cc: Youquan Song <youquan.song@intel.com>
Signed-off-by: Len Brown <len.brown@intel.com>
turbostat -s
cuts down on the amount of output, per user request.
also treak some output whitespace and the man page.
Signed-off-by: Len Brown <len.brown@intel.com>
This patch allows cpupower tool to generate its output files in a
seperate directory. This is now possible by passing the 'O=<path>' to
the command line.
This can be usefull for a normal user if the kernel source code is
located in a read only location.
This is patch stole some bits of the perf makefile.
[linux@dominikbrodowski.net: fix commit message]
Signed-off-by: Franck Bui-Huu <fbuihuu@gmail.com>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
Use the '-p' and '-o' switches to specify the pathname of the output
file to xgettext(1). This avoids to move manually the output file if
xgettext(1) succeeds.
Signed-off-by: Franck Bui-Huu <fbuihuu@gmail.com>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
UTIL_BINS and IDLE_OBJS variables are not defined at all, so
there's no need to remove their content from the 'clean' target.
Signed-off-by: Franck Bui-Huu <fbuihuu@gmail.com>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
Looks like some not needed debug code slipped in.
Also this code:
tmp = sysfs_get_idlestate_name(cpu, idlestates - 1);
performs a strdup and the mem was not freed again.
-> delete it.
Signed-off-by: Thomas Renninger <trenn@suse.de>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
The number of idle states was wrong.
The POLL idle state (on X86) was missed out:
Number of idle states: 4
Available idle states: C1-NHM C3-NHM C6-NHM
While the POLL is not a real idle state, its
statistics should still be shown. It's now also
explained in a detailed manpage.
This should fix a bug of missing the first idle
state on other archs.
Signed-off-by: Thomas Renninger <trenn@suse.de>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
cpupower-frequency-* manpages slightly differed from the others.
- Use uppercase letters in the title
- Show cpupower Manual in the header
- Remove Mattia from left down corner of the manpage, he is already
listed as author
- Remove --help, prints this message -> not needed
Signed-off-by: Thomas Renninger <trenn@suse.de>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
The last missing manpage for cpupower tools.
More info about other architecture's sleep state specialities would be great.
Signed-off-by: Thomas Renninger <trenn@suse.de>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
The name of the monitor is updated at runtime to the name of the
CPU type.
Signed-off-by: Thomas Renninger <trenn@suse.de>
CC: Andreas Herrmann <herrmann.der.user@googlemail.com>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
AMD's BKDG (Bios and Kernel Developers Guide) talks in the CPU spec of their
CPU families about PCI registers defined by "device" (slot) and func(tion).
Assuming that CPU specific configuration PCI devices are always on domain
and bus zero a pci_slot_func_init() func which gets the slot and func of
the desired PCI device passed looks like the most convenient way.
This also obsoletes the PCI device id maintenance.
Signed-off-by: Thomas Renninger <trenn@suse.de>
CC: Andreas Herrmann <herrmann.der.user@googlemail.com>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
This includes initial support for the recently published ACPI 5.0 spec.
In particular, support for the "hardware-reduced" bit that eliminates
the dependency on legacy hardware.
APEI has patches resulting from testing on real hardware.
Plus other random fixes.
* 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/lenb/linux: (52 commits)
acpi/apei/einj: Add extensions to EINJ from rev 5.0 of acpi spec
intel_idle: Split up and provide per CPU initialization func
ACPI processor: Remove unneeded variable passed by acpi_processor_hotadd_init V2
ACPI processor: Remove unneeded cpuidle_unregister_driver call
intel idle: Make idle driver more robust
intel_idle: Fix a cast to pointer from integer of different size warning in intel_idle
ACPI: kernel-parameters.txt : Add intel_idle.max_cstate
intel_idle: remove redundant local_irq_disable() call
ACPI processor: Fix error path, also remove sysdev link
ACPI: processor: fix acpi_get_cpuid for UP processor
intel_idle: fix API misuse
ACPI APEI: Convert atomicio routines
ACPI: Export interfaces for ioremapping/iounmapping ACPI registers
ACPI: Fix possible alignment issues with GAS 'address' references
ACPI, ia64: Use SRAT table rev to use 8bit or 16/32bit PXM fields (ia64)
ACPI, x86: Use SRAT table rev to use 8bit or 32bit PXM fields (x86/x86-64)
ACPI: Store SRAT table revision
ACPI, APEI, Resolve false conflict between ACPI NVS and APEI
ACPI, Record ACPI NVS regions
ACPI, APEI, EINJ, Refine the fix of resource conflict
...
Field names were shortened: "pkg" is now "pk", "core" is now "cr"
Signed-off-by: Arun Thomas <arun.thomas@gmail.com>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
Instead of printing something non-formatted to stdout, call
man(1) to show the man page for the proper subcommand.
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
Loosely based on a patch for cpufrequtils, submittted by
Sergey Dryabzhinsky <sergey.dryabzhinsky@gmail.com> and
signed-off-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
This allows for example:
cpupower -c 2-4,6 monitor -m Mperf
|Mperf
PKG |CORE|CPU | C0 | Cx | Freq
0| 8| 4| 2.42| 97.58| 1353
0| 16| 2| 14.38| 85.62| 1928
0| 24| 6| 1.76| 98.24| 1442
1| 16| 3| 15.53| 84.47| 1650
CPUs always get resorted for package, core then cpu id if it could get read out
(or however you name these topology levels...).
Still this is a nice way to keep the overview if a test binary is bound to
a specific CPU or if one wants to show all CPUs inside a package or similar.
Still missing: Do not measure not available cores to reduce the overhead
and achieve better results.
Signed-off-by: Thomas Renninger <trenn@suse.de>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
Before, checking for offlined CPUs was done dirty and
it was checked whether topology parsing returned -1 values.
But this is a valid case on a Xen (and possibly other) kernels.
Do proper online/offline checking, also take CONFIG_HOTPLUG_CPU
option into account (no /sys/devices/../cpuX/online file).
Signed-off-by: Thomas Renninger <trenn@suse.de>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
Which makes the implementation independent from cpufreq drivers.
Therefore this would also work on a Xen kernel where the hypervisor
is doing frequency switching and idle entering.
Signed-off-by: Thomas Renninger <trenn@suse.de>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
* 'tools-release' of git://git.kernel.org/pub/scm/linux/kernel/git/lenb/linux-idle-2.6:
tools/power turbostat: fit output into 80 columns on snb-ep
tools/power x86_energy_perf_policy: fix print of uninitialized string
Reduce columns for package number to 1.
If you can afford more than 9 packages,
you can also afford a terminal with more than 80 columns:-)
Also shave a column also off the package C-states
Signed-off-by: Len Brown <len.brown@intel.com>
IA32-Intel Devel guide Volume 3A - 14.3.2.1
-------------------------------------------
...
Opportunistic processor performance operation can be disabled by setting bit 38 of
IA32_MISC_ENABLES. This mechanism is intended for BIOS only. If
IA32_MISC_ENABLES[38] is set, CPUID.06H:EAX[1] will return 0.
Better detect things via cpuid, this cleans up the code a bit
and the MSR parts were not working correctly anyway.
Signed-off-by: Thomas Renninger <trenn@suse.de>
CC: lenb@kernel.org
CC: linux@dominikbrodowski.net
CC: cpufreq@vger.kernel.org
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
This adds the last piece missing from turbostat (if called with -v).
It shows on Intel machines supporting Turbo Boost how many cores
have to be active/idle to enter which boost mode (frequency).
Whether the HW really enters these boost modes can be verified via
./cpupower monitor.
Signed-off-by: Thomas Renninger <trenn@suse.de>
CC: lenb@kernel.org
CC: linux@dominikbrodowski.net
CC: cpufreq@vger.kernel.org
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
larger sysfs data (>255 bytes) was truncated and thus used improperly
[linux@dominikbrodowski.net: adapted to cpupowerutils]
Signed-off-by: Roman Vasiyarov <rvasiyarov@gmail.com>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
As cpupowerutils is intended to be included into the kernel sources,
use the kernel versioning instead of a custom version.
The script utils/version-gen.sh is largely based on the script already
found in tools/perf/util/PERF-VERSION-GEN .
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
Use the quiet/verbose mechanism found in kernel tools, without
relying on the special tool "ccdv"
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
CPU power consumption vs performance tuning is no longer
limited to CPU frequency switching anymore: deep sleep states,
traditional dynamic frequency scaling and hidden turbo/boost
frequencies are tied close together and depend on each other.
The first two exist on different architectures like PPC, Itanium and
ARM, the latter (so far) only on X86. On X86 the APU (CPU+GPU) will
only run most efficiently if CPU and GPU has proper power management
in place.
Users and Developers want to have *one* tool to get an overview what
their system supports and to monitor and debug CPU power management
in detail. The tool should compile and work on as many architectures
as possible.
Once this tool stabilizes a bit, it is intended to replace the
Intel-specific tools in tools/power/x86
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
Follow kernel coding style traditions more closely.
Delete typedef, re-name "per cpu counters" to
simply be counters etc.
This patch changes no functionality.
Suggested-by: Thiago Farina <tfransosi@gmail.com>
Signed-off-by: Len Brown <len.brown@intel.com>
bug could cause false positive on indicating
presence of invarient TSC or APERF support.
Signed-off-by: Thomas Renninger <trenn@suse.de>
Signed-off-by: Len Brown <len.brown@intel.com>
MSR_IA32_ENERGY_PERF_BIAS first became available on Westmere Xeon.
It is implemented in all Sandy Bridge processors -- mobile, desktop and server.
It is expected to become increasingly important in subsequent generations.
x86_energy_perf_policy is a user-space utility to set the
hardware energy vs performance policy hint in the processor.
Most systems would benefit from "x86_energy_perf_policy normal"
at system startup, as the hardware default is maximum performance
at the expense of energy efficiency.
See x86_energy_perf_policy.8 man page for more information.
Background:
Linux-2.6.36 added "epb" to /proc/cpuinfo to indicate
if an x86 processor supports MSR_IA32_ENERGY_PERF_BIAS,
without actually modifying the MSR.
In March, 2010, Venkatesh Pallipadi proposed a small driver
that programmed MSR_IA32_ENERGY_PERF_BIAS, based on
the cpufreq governor in use. It also offered
a boot-time cmdline option to override.
http://lkml.org/lkml/2010/3/4/457
But hiding the hardware policy behind the
governor choice was deemed "kinda icky".
In June, 2010, I proposed a generic user/kernel API to
generalize the power/performance policy trade-off.
"RFC: /sys/power/policy_preference"
http://lkml.org/lkml/2010/6/16/399
That is my preference for implementing this capability,
but I received no support on the list.
So in September, 2010, I sent x86_energy_perf_policy.c to LKML,
a user-space utility that scribbles directly to the MSR.
http://lkml.org/lkml/2010/9/28/246
Here is that same utility, after responding to some review feedback,
to live in tools/power/, where it is easily found.
Signed-off-by: Len Brown <len.brown@intel.com>
turbostat is a Linux tool to observe proper operation
of Intel(R) Turbo Boost Technology.
turbostat displays the actual processor frequency
on x86 processors that include APERF and MPERF MSRs.
Note that turbostat is of limited utility on Linux
kernels 2.6.29 and older, as acpi_cpufreq cleared
APERF/MPERF up through that release.
On Intel Core i3/i5/i7 (Nehalem) and newer processors,
turbostat also displays residency in idle power saving states,
which are necessary for diagnosing any cpuidle issues
that may have an effect on turbo-mode.
See the turbostat.8 man page for example usage.
Signed-off-by: Len Brown <len.brown@intel.com>
