Revert commit 69a37bea (cpuidle: Quickly notice prediction failure for
repeat mode), because it has been identified as the source of a
significant performance regression in v3.8 and later as explained by
Jeremy Eder:
We believe we've identified a particular commit to the cpuidle code
that seems to be impacting performance of variety of workloads.
The simplest way to reproduce is using netperf TCP_RR test, so
we're using that, on a pair of Sandy Bridge based servers. We also
have data from a large database setup where performance is also
measurably/positively impacted, though that test data isn't easily
share-able.
Included below are test results from 3 test kernels:
kernel reverts
-----------------------------------------------------------
1) vanilla upstream (no reverts)
2) perfteam2 reverts e11538d1f03914eb92af5a1a378375c05ae8520c
3) test reverts 69a37beabf1f0a6705c08e879bdd5d82ff6486c4
e11538d1f03914eb92af5a1a378375c05ae8520c
In summary, netperf TCP_RR numbers improve by approximately 4%
after reverting 69a37beabf1f0a6705c08e879bdd5d82ff6486c4. When
69a37beabf1f0a6705c08e879bdd5d82ff6486c4 is included, C0 residency
never seems to get above 40%. Taking that patch out gets C0 near
100% quite often, and performance increases.
The below data are histograms representing the %c0 residency @
1-second sample rates (using turbostat), while under netperf test.
- If you look at the first 4 histograms, you can see %c0 residency
almost entirely in the 30,40% bin.
- The last pair, which reverts 69a37beabf1f0a6705c08e879bdd5d82ff6486c4,
shows %c0 in the 80,90,100% bins.
Below each kernel name are netperf TCP_RR trans/s numbers for the
particular kernel that can be disclosed publicly, comparing the 3
test kernels. We ran a 4th test with the vanilla kernel where
we've also set /dev/cpu_dma_latency=0 to show overall impact
boosting single-threaded TCP_RR performance over 11% above
baseline.
3.10-rc2 vanilla RX + c0 lock (/dev/cpu_dma_latency=0):
TCP_RR trans/s 54323.78
-----------------------------------------------------------
3.10-rc2 vanilla RX (no reverts)
TCP_RR trans/s 48192.47
Receiver %c0
0.0000 - 10.0000 [ 1]: *
10.0000 - 20.0000 [ 0]:
20.0000 - 30.0000 [ 0]:
30.0000 - 40.0000 [ 59]:
***********************************************************
40.0000 - 50.0000 [ 1]: *
50.0000 - 60.0000 [ 0]:
60.0000 - 70.0000 [ 0]:
70.0000 - 80.0000 [ 0]:
80.0000 - 90.0000 [ 0]:
90.0000 - 100.0000 [ 0]:
Sender %c0
0.0000 - 10.0000 [ 1]: *
10.0000 - 20.0000 [ 0]:
20.0000 - 30.0000 [ 0]:
30.0000 - 40.0000 [ 11]: ***********
40.0000 - 50.0000 [ 49]:
*************************************************
50.0000 - 60.0000 [ 0]:
60.0000 - 70.0000 [ 0]:
70.0000 - 80.0000 [ 0]:
80.0000 - 90.0000 [ 0]:
90.0000 - 100.0000 [ 0]:
-----------------------------------------------------------
3.10-rc2 perfteam2 RX (reverts commit
e11538d1f03914eb92af5a1a378375c05ae8520c)
TCP_RR trans/s 49698.69
Receiver %c0
0.0000 - 10.0000 [ 1]: *
10.0000 - 20.0000 [ 1]: *
20.0000 - 30.0000 [ 0]:
30.0000 - 40.0000 [ 59]:
***********************************************************
40.0000 - 50.0000 [ 0]:
50.0000 - 60.0000 [ 0]:
60.0000 - 70.0000 [ 0]:
70.0000 - 80.0000 [ 0]:
80.0000 - 90.0000 [ 0]:
90.0000 - 100.0000 [ 0]:
Sender %c0
0.0000 - 10.0000 [ 1]: *
10.0000 - 20.0000 [ 0]:
20.0000 - 30.0000 [ 0]:
30.0000 - 40.0000 [ 2]: **
40.0000 - 50.0000 [ 58]:
**********************************************************
50.0000 - 60.0000 [ 0]:
60.0000 - 70.0000 [ 0]:
70.0000 - 80.0000 [ 0]:
80.0000 - 90.0000 [ 0]:
90.0000 - 100.0000 [ 0]:
-----------------------------------------------------------
3.10-rc2 test RX (reverts 69a37beabf1f0a6705c08e879bdd5d82ff6486c4
and e11538d1f03914eb92af5a1a378375c05ae8520c)
TCP_RR trans/s 47766.95
Receiver %c0
0.0000 - 10.0000 [ 1]: *
10.0000 - 20.0000 [ 1]: *
20.0000 - 30.0000 [ 0]:
30.0000 - 40.0000 [ 27]: ***************************
40.0000 - 50.0000 [ 2]: **
50.0000 - 60.0000 [ 0]:
60.0000 - 70.0000 [ 2]: **
70.0000 - 80.0000 [ 0]:
80.0000 - 90.0000 [ 0]:
90.0000 - 100.0000 [ 28]: ****************************
Sender:
0.0000 - 10.0000 [ 1]: *
10.0000 - 20.0000 [ 0]:
20.0000 - 30.0000 [ 0]:
30.0000 - 40.0000 [ 11]: ***********
40.0000 - 50.0000 [ 0]:
50.0000 - 60.0000 [ 1]: *
60.0000 - 70.0000 [ 0]:
70.0000 - 80.0000 [ 3]: ***
80.0000 - 90.0000 [ 7]: *******
90.0000 - 100.0000 [ 38]: **************************************
These results demonstrate gaining back the tendency of the CPU to
stay in more responsive, performant C-states (and thus yield
measurably better performance), by reverting commit
69a37beabf1f0a6705c08e879bdd5d82ff6486c4.
Requested-by: Jeremy Eder <jeder@redhat.com>
Tested-by: Len Brown <len.brown@intel.com>
Cc: 3.8+ <stable@vger.kernel.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Revert commit e11538d1 (cpuidle: Quickly notice prediction failure in general case), since it depends on commit 69a37be (cpuidle: Quickly notice prediction failure for repeat mode) that has been identified as the source of a significant performance regression in v3.8 and later. Requested-by: Jeremy Eder <jeder@redhat.com> Tested-by: Len Brown <len.brown@intel.com> Cc: 3.8+ <stable@vger.kernel.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Pull power management and ACPI updates from Rafael Wysocki:
"This time the total number of ACPI commits is slightly greater than
the number of cpufreq commits, but Viresh Kumar (who works on cpufreq)
remains the most active patch submitter.
To me, the most significant change is the addition of offline/online
device operations to the driver core (with the Greg's blessing) and
the related modifications of the ACPI core hotplug code. Next are the
freezer updates from Colin Cross that should make the freezing of
tasks a bit less heavy weight.
We also have a couple of regression fixes, a number of fixes for
issues that have not been identified as regressions, two new drivers
and a bunch of cleanups all over.
Highlights:
- Hotplug changes to support graceful hot-removal failures.
It sometimes is necessary to fail device hot-removal operations
gracefully if they cannot be carried out completely. For example,
if memory from a memory module being hot-removed has been allocated
for the kernel's own use and cannot be moved elsewhere, it's
desirable to fail the hot-removal operation in a graceful way
rather than to crash the kernel, but currenty a success or a kernel
crash are the only possible outcomes of an attempted memory
hot-removal. Needless to say, that is not a very attractive
alternative and it had to be addressed.
However, in order to make it work for memory, I first had to make
it work for CPUs and for this purpose I needed to modify the ACPI
processor driver. It's been split into two parts, a resident one
handling the low-level initialization/cleanup and a modular one
playing the actual driver's role (but it binds to the CPU system
device objects rather than to the ACPI device objects representing
processors). That's been sort of like a live brain surgery on a
patient who's riding a bike.
So this is a little scary, but since we found and fixed a couple of
regressions it caused to happen during the early linux-next testing
(a month ago), nobody has complained.
As a bonus we remove some duplicated ACPI hotplug code, because the
ACPI-based CPU hotplug is now going to use the common ACPI hotplug
code.
- Lighter weight freezing of tasks.
These changes from Colin Cross and Mandeep Singh Baines are
targeted at making the freezing of tasks a bit less heavy weight
operation. They reduce the number of tasks woken up every time
during the freezing, by using the observation that the freezer
simply doesn't need to wake up some of them and wait for them all
to call refrigerator(). The time needed for the freezer to decide
to report a failure is reduced too.
Also reintroduced is the check causing a lockdep warining to
trigger when try_to_freeze() is called with locks held (which is
generally unsafe and shouldn't happen).
- cpufreq updates
First off, a commit from Srivatsa S Bhat fixes a resume regression
introduced during the 3.10 cycle causing some cpufreq sysfs
attributes to return wrong values to user space after resume. The
fix is kind of fresh, but also it's pretty obvious once Srivatsa
has identified the root cause.
Second, we have a new freqdomain_cpus sysfs attribute for the
acpi-cpufreq driver to provide information previously available via
related_cpus. From Lan Tianyu.
Finally, we fix a number of issues, mostly related to the
CPUFREQ_POSTCHANGE notifier and cpufreq Kconfig options and clean
up some code. The majority of changes from Viresh Kumar with bits
from Jacob Shin, Heiko Stübner, Xiaoguang Chen, Ezequiel Garcia,
Arnd Bergmann, and Tang Yuantian.
- ACPICA update
A usual bunch of updates from the ACPICA upstream.
During the 3.4 cycle we introduced support for ACPI 5 extended
sleep registers, but they are only supposed to be used if the
HW-reduced mode bit is set in the FADT flags and the code attempted
to use them without checking that bit. That caused suspend/resume
regressions to happen on some systems. Fix from Lv Zheng causes
those registers to be used only if the HW-reduced mode bit is set.
Apart from this some other ACPICA bugs are fixed and code cleanups
are made by Bob Moore, Tomasz Nowicki, Lv Zheng, Chao Guan, and
Zhang Rui.
- cpuidle updates
New driver for Xilinx Zynq processors is added by Michal Simek.
Multidriver support simplification, addition of some missing
kerneldoc comments and Kconfig-related fixes come from Daniel
Lezcano.
- ACPI power management updates
Changes to make suspend/resume work correctly in Xen guests from
Konrad Rzeszutek Wilk, sparse warning fix from Fengguang Wu and
cleanups and fixes of the ACPI device power state selection
routine.
- ACPI documentation updates
Some previously missing pieces of ACPI documentation are added by
Lv Zheng and Aaron Lu (hopefully, that will help people to
uderstand how the ACPI subsystem works) and one outdated doc is
updated by Hanjun Guo.
- Assorted ACPI updates
We finally nailed down the IA-64 issue that was the reason for
reverting commit 9f29ab11ddbf ("ACPI / scan: do not match drivers
against objects having scan handlers"), so we can fix it and move
the ACPI scan handler check added to the ACPI video driver back to
the core.
A mechanism for adding CMOS RTC address space handlers is
introduced by Lan Tianyu to allow some EC-related breakage to be
fixed on some systems.
A spec-compliant implementation of acpi_os_get_timer() is added by
Mika Westerberg.
The evaluation of _STA is added to do_acpi_find_child() to avoid
situations in which a pointer to a disabled device object is
returned instead of an enabled one with the same _ADR value. From
Jeff Wu.
Intel BayTrail PCH (Platform Controller Hub) support is added to
the ACPI driver for Intel Low-Power Subsystems (LPSS) and that
driver is modified to work around a couple of known BIOS issues.
Changes from Mika Westerberg and Heikki Krogerus.
The EC driver is fixed by Vasiliy Kulikov to use get_user() and
put_user() instead of dereferencing user space pointers blindly.
Code cleanups are made by Bjorn Helgaas, Nicholas Mazzuca and Toshi
Kani.
- Assorted power management updates
The "runtime idle" helper routine is changed to take the return
values of the callbacks executed by it into account and to call
rpm_suspend() if they return 0, which allows us to reduce the
overall code bloat a bit (by dropping some code that's not
necessary any more after that modification).
The runtime PM documentation is updated by Alan Stern (to reflect
the "runtime idle" behavior change).
New trace points for PM QoS are added by Sahara
(<keun-o.park@windriver.com>).
PM QoS documentation is updated by Lan Tianyu.
Code cleanups are made and minor issues are addressed by Bernie
Thompson, Bjorn Helgaas, Julius Werner, and Shuah Khan.
- devfreq updates
New driver for the Exynos5-bus device from Abhilash Kesavan.
Minor cleanups, fixes and MAINTAINERS update from MyungJoo Ham,
Abhilash Kesavan, Paul Bolle, Rajagopal Venkat, and Wei Yongjun.
- OMAP power management updates
Adaptive Voltage Scaling (AVS) SmartReflex voltage control driver
updates from Andrii Tseglytskyi and Nishanth Menon."
* tag 'pm+acpi-3.11-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm: (162 commits)
cpufreq: Fix cpufreq regression after suspend/resume
ACPI / PM: Fix possible NULL pointer deref in acpi_pm_device_sleep_state()
PM / Sleep: Warn about system time after resume with pm_trace
cpufreq: don't leave stale policy pointer in cdbs->cur_policy
acpi-cpufreq: Add new sysfs attribute freqdomain_cpus
cpufreq: make sure frequency transitions are serialized
ACPI: implement acpi_os_get_timer() according the spec
ACPI / EC: Add HP Folio 13 to ec_dmi_table in order to skip DSDT scan
ACPI: Add CMOS RTC Operation Region handler support
ACPI / processor: Drop unused variable from processor_perflib.c
cpufreq: tegra: call CPUFREQ_POSTCHANGE notfier in error cases
cpufreq: s3c64xx: call CPUFREQ_POSTCHANGE notfier in error cases
cpufreq: omap: call CPUFREQ_POSTCHANGE notfier in error cases
cpufreq: imx6q: call CPUFREQ_POSTCHANGE notfier in error cases
cpufreq: exynos: call CPUFREQ_POSTCHANGE notfier in error cases
cpufreq: dbx500: call CPUFREQ_POSTCHANGE notfier in error cases
cpufreq: davinci: call CPUFREQ_POSTCHANGE notfier in error cases
cpufreq: arm-big-little: call CPUFREQ_POSTCHANGE notfier in error cases
cpufreq: powernow-k8: call CPUFREQ_POSTCHANGE notfier in error cases
cpufreq: pcc: call CPUFREQ_POSTCHANGE notfier in error cases
...
Pull ARM SoC specific changes from Arnd Bergmann: "These changes are all to SoC-specific code, a total of 33 branches on 17 platforms were pulled into this. Like last time, Renesas sh-mobile is now the platform with the most changes, followed by OMAP and EXYNOS. Two new platforms, TI Keystone and Rockchips RK3xxx are added in this branch, both containing almost no platform specific code at all, since they are using generic subsystem interfaces for clocks, pinctrl, interrupts etc. The device drivers are getting merged through the respective subsystem maintainer trees. One more SoC (u300) is now multiplatform capable and several others (shmobile, exynos, msm, integrator, kirkwood, clps711x) are moving towards that goal with this series but need more work. Also noteworthy is the work on PCI here, which is traditionally part of the SoC specific code. With the changes done by Thomas Petazzoni, we can now more easily have PCI host controller drivers as loadable modules and keep them separate from the platform code in drivers/pci/host. This has already led to the discovery that three platforms (exynos, spear and imx) are actually using an identical PCIe host controller and will be able to share a driver once support for spear and imx is added." * tag 'soc-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/arm/arm-soc: (480 commits) ARM: integrator: let pciv3 use mem/premem from device tree ARM: integrator: set local side PCI addresses right ARM: dts: Add pcie controller node for exynos5440-ssdk5440 ARM: dts: Add pcie controller node for Samsung EXYNOS5440 SoC ARM: EXYNOS: Enable PCIe support for Exynos5440 pci: Add PCIe driver for Samsung Exynos ARM: OMAP5: voltagedomain data: remove temporary OMAP4 voltage data ARM: keystone: Move CPU bringup code to dedicated asm file ARM: multiplatform: always pick one CPU type ARM: imx: select syscon for IMX6SL ARM: keystone: select ARM_ERRATA_798181 only for SMP ARM: imx: Synertronixx scb9328 needs to select SOC_IMX1 ARM: OMAP2+: AM43x: resolve SMP related build error dmaengine: edma: enable build for AM33XX ARM: edma: Add EDMA crossbar event mux support ARM: edma: Add DT and runtime PM support to the private EDMA API dmaengine: edma: Add TI EDMA device tree binding arm: add basic support for Rockchip RK3066a boards arm: add debug uarts for rockchip rk29xx and rk3xxx series arm: Add basic clocks for Rockchip rk3066a SoCs ...
Like other ARM specific drivers, this one requires ARM_CPU_SUSPEND, as shown by this linker error: drivers/built-in.o: In function `calxeda_pwrdown_idle': drivers/cpuidle/cpuidle-calxeda.c:84: undefined reference to `cpu_suspend' drivers/cpuidle/cpuidle-calxeda.c:86: undefined reference to `cpu_resume' Signed-off-by: Arnd Bergmann <arnd@arndb.de> Acked-by: Rafael J. Wysocki <rjw@sisk.pl> Acked-by: Rob Herring <rob.herring@calxeda.com> Acked-by: Daniel Lezcano <daniel.lezcano@linaro.org> Cc: linux-pm@vger.kernel.org
Before commit d6f346f (cpuidle: improve governor Kconfig options), the CONFIG_ARCH_NEEDS_CPU_IDLE_COUPLED option didn't depend on CONFIG_CPU_IDLE but now it has been moved under the CPU_IDLE menuconfig. That raises the following warnings: warning: (ARCH_OMAP4 && ARCH_TEGRA_2x_SOC) selects ARCH_NEEDS_CPU_IDLE_COUPLED which has unmet direct dependencies (CPU_IDLE) warning: (ARCH_OMAP4 && ARCH_TEGRA_2x_SOC) selects ARCH_NEEDS_CPU_IDLE_COUPLED which has unmet direct dependencies (CPU_IDLE) because the tegra2 and omap4 Kconfig files select this option without checking if CPU_IDLE is set. Fix that by moving ARCH_NEEDS_CPU_IDLE_COUPLED outside of CPU_IDLE. [rjw: Changelog] Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Add kerneldoc (and other) comments to the cpuidle driver's framework code. Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Commit bf4d1b5 (cpuidle: support multiple drivers) introduced support for using multiple cpuidle drivers at the same time. It added a couple of new APIs to register the driver per CPU, but that led to some unnecessary code complexity related to the kernel config options deciding whether or not the multiple driver support is enabled. The code has to work as it did before when the multiple driver support is not enabled and the multiple driver support has to be compatible with the previously existing API. Remove the new API, not used by any driver in the tree yet (but needed for the HMP cpuidle drivers that will be submitted soon), and add a new cpumask pointer to the cpuidle driver structure that will point to the mask of CPUs handled by the given driver. That will allow the cpuidle_[un]register_driver() API to be used for the multiple driver support along with the cpuidle_[un]register() functions added recently. [rjw: Changelog] Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Add cpuidle support for Xilinx Zynq. Signed-off-by: Michal Simek <michal.simek@xilinx.com> Acked-by: Daniel Lezcano <daniel.lezcano@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Each governor is suitable for different kernel configurations: the menu governor suits better for a tickless system, while the ladder governor fits better for a periodic timer tick system. The Kconfig does not allow to [un]select a governor, thus both are compiled in the kernel but the init order makes the menu governor to be the last one to be registered, so becoming the default. The only way to switch back to the ladder governor is to enable the sysfs governor switch in the kernel command line. Because it seems nobody complained about this, the menu governor is used by default most of the time on the system, having both governors is not really necessary on a tickless system but there isn't a config option to disable one or another governor. Create a submenu for cpuidle and add a label for each governor, so we can see the option in the menu config and enable/disable it. The governors will be enabled depending on the CONFIG_NO_HZ option: - If CONFIG_NO_HZ is set, then the menu governor is selected and the ladder governor is optional, defaulting to 'yes' - If CONFIG_NO_HZ is not set, then the ladder governor is selected and the menu governor is optional, defaulting to 'yes' Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>