The original intent here was to track unweighted runqueue load
with less resolution so we could use the least-recently-disturbed
runqueue to choose between 'closely related' load levels.
However, after experimenting with the resolution it turns out
that the following algorithm is highly beneficial for mobile
workloads.
In hmp_domain_min_load:
* If any CPU is zero, the overall load is zero
* If no CPUs are idle, the domain is 'fully loaded'
Additionally, the time since last migration count is used to
discriminate between idle CPUs.
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
Signed-off-by: Liviu Dudau <liviu.dudau@arm.com>
Signed-off-by: Jon Medhurst <tixy@linaro.org>
Track when migrations were performed to runqueues.
Use this to decide between runqueues as migration targets when run
queues in an hmp domain have equal load.
Intention is to spread migration load amongst CPUs more fairly.
When all CPUs in an hmp domain are fully loaded, the existing code
always selects the last CPU as a migration target - this is unfair
and little better than doing no selection.
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
Signed-off-by: Liviu Dudau <liviu.dudau@arm.com>
Signed-off-by: Jon Medhurst <tixy@linaro.org>
The hmp_get_{lightest,heaviest}_task() need to use
__pick_first_entity() to get a pointer to a sched_entity on the rq.
The current is not kept on the rq while running, so its rb-tree node
pointers are no longer valid.
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
Signed-off-by: Liviu Dudau <liviu.dudau@arm.com>
Signed-off-by: Jon Medhurst <tixy@linaro.org>
When we are looking for a task to migrate up, select the heaviest
one in the first 5 runnable on the runqueue.
Likewise, when looking for a task to offload, select the lightest
one in the first 5 runnable on the runqueue.
Ensure task selected is runnable in the target domain.
This change is necessary in order to implement idle pull in a
sensible manner, but here is used in up-migration and offload to
select the correct target task.
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
Signed-off-by: Liviu Dudau <liviu.dudau@arm.com>
Signed-off-by: Jon Medhurst <tixy@linaro.org>
It is sometimes desirable to run a kernel with HMP scheduling enabled
on a system which is not big.LITTLE, e.g. when building a multi-platform
kernel, or when testing a big.LITTLE system with one cluster disabled.
We should therefore allow for the situation where is no little domain.
Signed-off-by: Jon Medhurst <tixy@linaro.org>
Signed-off-by: Mark Brown <broonie@linaro.org>
This patch restricts the allowed cpu mask for rt tasks initially started
with a full cpu mask to the little domain.
An rt task is specified as real time in __setscheduler() which is finally
called for all rt tasks (kernel and user land). In this function we
restrict the allowed cpu mask to the little domain.
This also prevents that a rt tasks can later be pushed to the big domain
because the function find_lowest_rq() will only recognize the allowed cpu
mask of a task to find the new cpu the task runs on.
Current kludges of the patch:
* Since we do not have an API to get the cpu mask of the A7 cluster,
hmp_slow_cpu_mask is made global in arm/kernel/topology.c for now.
* The watchdog_enable() function calls sched_setscheduler() before
kthread_bind() for the cpu specific watchdog kernel threads. The order of
these two calls has to be changed to make this patch work.
Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
There is little point in doing a nohz balance kick on a CPU from a
different HMP domain, since the unset SD_LOAD_BALANCE flag on the CPU
domain level prevents tasks from being balanced across clusters
except through the per-task load driven hmp_migrate/hmp_offload paths.
Further, the nohz balance kick is actively harmful to power usage if
all the tasks fit into the little domain since it causes the big
domain to wake up and do a lot of calculation to determine that
there is nothing to do.
A more generic solution is to walk the sched domain tree and determine
the intersection of potential idle balance cpus with visibility of
tasks on the current CPU, however HMP domains are more easily
accessible.
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
Initialise the load stats for new tasks so that they do not
see the instability in early task life which makes it so hard to
decide which CPU is appropriate.
Also, change the fork balance algorithm so that the least loaded of
the CPUs in the big cluster is chosen regardless of the bigness of
the parent task.
This is intended to help performance for applications which use
many short-lived tasks. Although best practise is usually to use
a thread pool, apps which do not do this should not be subject to
the randomness of the early stats.
We should ignore real-time threads for forking on big CPUs, but
it is not possible to figure out if a new thread is real-time or
not at the fork stage. Instead, we prevent kernel threads from
getting the initial boost - when they later become real-time they
will only be on big if their compute requirements demand it.
Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
When evaluating a migration we make two calls to hmp_domain_min_load.
This is unnecessary if we pass on the target CPU information from the
hmp_up_migration path.
In hmp_down_migration, we don't consider the load of the target CPUS.
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
The reference patch set always selects the first CPU in an HMP
domain as a migration target. In busy situations, this means that
the migrated thread cannot make immediate use of an idle CPU but
must share a busy one until the load balancer runs across the big
domain.
This patch uses the hmp_domain_min_load function introduced in
global balancing to figure out which of the CPUs is the least busy
and selects that as a migration target - in both directions.
This essentially implements a task-spread strategy and is intended
to maximise performance of migrated threads but is likely
to use more power than the packing strategy previously employed.
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
Normal task and runqueue loading is scaled according to priority
to end up with a weighted load, known as the contribution.
We want the CPU time to be allotted according to priority, but
we also want to make big/little decisions based upon raw load.
It is common, for example, for Android apps following the dev
guide to end up with all their long-running or async action
threads as low priority unless they override the AsyncThread
constructor. All these threads are such low priority that they
become invisible to the hmp_offload routine.
Using unweighted load here allows us to maximise CPU usage in busy
situations.
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
This reverts commit 68315334e32932739145ddb41a46cc86b8b056b3.
Having the priority filter enabled prevents proper operation
on Android systems where a wider range of priorities are used
by userspace to partition types of tasks. Those tasks should still
be able to benefit from the use of big CPUs when required.
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
rq->nr_running is the actual number of runnable tasks we wish to use
to determine if a task is alone on a CPU.
Change-Id: Icaf3022e02924ecdc94e14d4146c6fadd9580e2b
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
This patch introduces an extra-check at task up-migration to
prevent overloading the cpus in the faster hmp_domain while the
slower hmp_domain is not fully utilized. The patch also introduces
a periodic balance check that can down-migrate tasks if the faster
domain is oversubscribed and the slower is under-utilized.
Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com>
Evaluation Patch to investigate using load as a representation of the
amount of POTENTIAL cpu compute capacity used rather than a representation
of the CURRENT cpu compute capacity.
If CPUFreq is enabled, scales load in accordance with frequency.
Powersave/performance CPUFreq governors are detected and scaling is
disabled while these governors are in use. This is because when a
single-frequency governor is in use, potential CPU capacity is static.
So long as the governors and CPUFreq subsystem correctly report the
frequencies available, the scaling should self tune.
Adds an additional file to sysfs to allow this feature to be disabled
for experimentation.
/sys/kernel/hmp/frequency_invariant_load_scale
write 0 to disable, 1 to enable.
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
These functions allow to change the load average period used
in the task load average computation through
/sys/kernel/hmp/load_avg_period_ms. This period is the time
in ms to go from 0 to 0.5 load average while running or the
time from 1 to 0.5 while sleeping.
The default one used is 32 and gives the same load_avg_ratio
computation than without this patch. These functions also allow
to change the up and down threshold of HMP using
/sys/kernel/hmp/{up,down}_threshold. Both must be between 0 and
1024. The thresholds are divided by 1024 before being compared
to the load_avg_ratio.
If /sys/kernel/hmp/load_avg_period_ms is 128 and
/sys/kernel/hmp/up_threshold is 512, a task will be migrated
to a bigger cluster after running for 128ms. Because after
load_avg_period_ms the load average is 0.5 and real up_threshold
us 512 / 1024 = 0.5.
Signed-off-by: Olivier Cozette <olivier.cozette@arm.com>
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
Previously, an offline CPU would always appear to have a zero load
and this would distort the offload functionality used for balancing
big and little domains.
Maintain a mask of online CPUs in each domain and use this instead.
Change-Id: I639b564b2f40cb659af8ceb8bd37f84b8a1fe323
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
If the entity is not a task, it is a cfs group rq. Iterate up to
find the task entity.
Change-Id: I7cab7aba0798f6f14e38ad32e566d90e5937ffbc
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
The patch "sched: Use device-tree to provide fast/slow CPU list for HMP"
depends on the ordering of CPU's in the device tree. It breaks to determine
the logical mask correctly if the logical mask of the CPUs differ from
physical ordering in the device tree.
This patch fix the logic by depending on the mpidr in the device tree
and mapping that mpidr to the logical cpu.
Signed-off-by: Sudeep KarkadaNagesha <sudeep.karkadanagesha@arm.com>
Signed-off-by: Liviu Dudau <Liviu.Dudau@arm.com>
Adds an extra check intersection of the task affinity mask and the slower
hmp_domain cpumask before down migrating low priority tasks.
Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com>
On homogeneous (non-heterogeneous) systems all CPUs will be declared
'fast' and the slow cpu list will be empty. In this situation we need to
avoid adding an empty slow HMP domain otherwise the scheduler code will
blow up when it attempts to move a task to the slow domain.
Signed-off-by: Jon Medhurst <tixy@linaro.org>
We need a way to prevent tasks that are migrating up and down the
hmp_domains from migrating straight on through before the load has
adapted to the new compute capacity of the CPU on the new hmp_domain.
This patch adds a next up/down migration delay that prevents the task
from doing another migration in the same direction until the delay
has expired.
Signed-off-by: Morten Rasmussen <Morten.Rasmussen@arm.com>
Adds ftrace events for key variables related to the entity
load-tracking to help debugging scheduler behaviour. Allows tracing
of load contribution and runqueue residency ratio for both entities
and runqueues as well as entity CPU usage ratio.
Signed-off-by: Morten Rasmussen <Morten.Rasmussen@arm.com>
SCHED_HMP requires the different cpu types to be represented by an
ordered list of hmp_domains. Each hmp_domain represents all cpus of
a particular type using a cpumask.
The list is platform specific and therefore must be generated by
platform code by implementing arch_get_hmp_domains().
Signed-off-by: Morten Rasmussen <Morten.Rasmussen@arm.com>
We can't rely on Kconfig options to set the fast and slow CPU lists for
HMP scheduling if we want a single kernel binary to support multiple
devices with different CPU topology. E.g. TC2 (ARM's Test-Chip-2
big.LITTLE system), Fast Models, or even non big.LITTLE devices.
This patch adds the function arch_get_fast_and_slow_cpus() to generate
the lists at run-time by parsing the CPU nodes in device-tree; it
assumes slow cores are A7s and everything else is fast. The function
still supports the old Kconfig options as this is useful for testing the
HMP scheduler on devices without big.LITTLE.
This patch is reuse of a patch by Jon Medhurst <tixy@linaro.org> with a
few bits left out.
Signed-off-by: Morten Rasmussen <Morten.Rasmussen@arm.com>
Adds Kconfig entries to enable HMP scheduling on ARM platforms.
Currently, it disables CPU level sched_domain load-balacing in order
to simplify things. This needs fixing in a later revision. HMP
scheduling will do the load-balancing at this level instead.
Signed-off-by: Morten Rasmussen <Morten.Rasmussen@arm.com>
Introduces a priority threshold which prevents low priority task
from migrating to faster hmp_domains (cpus). This is useful for
user-space software which assigns lower task priority to background
task.
Signed-off-by: Morten Rasmussen <Morten.Rasmussen@arm.com>
This patch introduces forced task migration for moving suitable
currently running tasks between hmp_domains. Task behaviour is likely
to change over time. Tasks running in a less capable hmp_domain may
change to become more demanding and should therefore be migrated up.
They are unlikely go through the select_task_rq_fair() path anytime
soon and therefore need special attention.
This patch introduces a period check (SCHED_TICK) of the currently
running task on all runqueues and sets up a forced migration using
stop_machine_no_wait() if the task needs to be migrated.
Ideally, this should not be implemented by polling all runqueues.
Signed-off-by: Morten Rasmussen <Morten.Rasmussen@arm.com>
This patch introduces the basic SCHED_HMP infrastructure. Each class of
cpus is represented by a hmp_domain and tasks will only be moved between
these domains when their load profiles suggest it is beneficial.
SCHED_HMP relies heavily on the task load-tracking introduced in Paul
Turners fair group scheduling patch set:
<https://lkml.org/lkml/2012/8/23/267>
SCHED_HMP requires that the platform implements arch_get_hmp_domains()
which should set up the platform specific list of hmp_domains. It is
also assumed that the platform disables SD_LOAD_BALANCE for the
appropriate sched_domains.
Tasks placement takes place every time a task is to be inserted into
a runqueue based on its load history. The task placement decision is
based on load thresholds.
There are no restrictions on the number of hmp_domains, however,
multiple (>2) has not been tested and the up/down migration policy is
rather simple.
Signed-off-by: Morten Rasmussen <Morten.Rasmussen@arm.com>
This patch adds load_avg_ratio to each task. The load_avg_ratio is a
variant of load_avg_contrib which is not scaled by the task priority. It
is calculated like this:
runnable_avg_sum * NICE_0_LOAD / (runnable_avg_period + 1).
Signed-off-by: Morten Rasmussen <Morten.Rasmussen@arm.com>
With the frame-work for runnable tracking now fully in place. Per-entity usage
tracking is a simple and low-overhead addition.
Signed-off-by: Paul Turner <pjt@google.com>
Reviewed-by: Ben Segall <bsegall@google.com>
Pull another powerpc fix from Benjamin Herrenschmidt:
"I mentioned that while we had fixed the kernel crashes, EEH error
recovery didn't always recover... It appears that I had a fix for
that already in powerpc-next (with a stable CC).
I cherry-picked it today and did a few tests and it seems that things
now work quite well. The patch is also pretty simple, so I see no
reason to wait before merging it."
* 'merge' of git://git.kernel.org/pub/scm/linux/kernel/git/benh/powerpc:
powerpc/eeh: Fix fetching bus for single-dev-PE
Pull SCSI fixes from James Bottomley:
"This is a set of seven bug fixes. Several fcoe fixes for locking
problems, initiator issues and a VLAN API change, all of which could
eventually lead to data corruption, one fix for a qla2xxx locking
problem which could lead to multiple completions of the same request
(and subsequent data corruption) and a use after free in the ipr
driver. Plus one minor MAINTAINERS file update"
(only six bugfixes in this pull, since I had already pulled the fcoe API
fix directly from Robert Love)
* tag 'scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi:
[SCSI] ipr: Avoid target_destroy accessing memory after it was freed
[SCSI] qla2xxx: Fix for locking issue between driver ISR and mailbox routines
MAINTAINERS: Fix fcoe mailing list
libfc: extend ex_lock to protect all of fc_seq_send
libfc: Correct check for initiator role
libfcoe: Fix Conflicting FCFs issue in the fabric
While running Linux as guest on top of phyp, we possiblly have
PE that includes single PCI device. However, we didn't return
its PCI bus correctly and it leads to failure on recovery from
EEH errors for single-dev-PE. The patch fixes the issue.
Cc: <stable@vger.kernel.org> # v3.7+
Cc: Steve Best <sbest@us.ibm.com>
Signed-off-by: Gavin Shan <shangw@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Pull powerpc fixes from Ben Herrenschmidt:
"We discovered some breakage in our "EEH" (PCI Error Handling) code
while doing error injection, due to a couple of regressions. One of
them is due to a patch (37f02195be "powerpc/pci: fix PCI-e devices
rescan issue on powerpc platform") that, in hindsight, I shouldn't
have merged considering that it caused more problems than it solved.
Please pull those two fixes. One for a simple EEH address cache
initialization issue. The other one is a patch from Guenter that I
had originally planned to put in 3.11 but which happens to also fix
that other regression (a kernel oops during EEH error handling and
possibly hotplug).
With those two, the couple of test machines I've hammered with error
injection are remaining up now. EEH appears to still fail to recover
on some devices, so there is another problem that Gavin is looking
into but at least it's no longer crashing the kernel."
* 'merge' of git://git.kernel.org/pub/scm/linux/kernel/git/benh/powerpc:
powerpc/pci: Improve device hotplug initialization
powerpc/eeh: Add eeh_dev to the cache during boot
Due to recent changes and expecations of proper cpu bindings, there are
now cases for many of the in-tree devicetrees where a WARN() will hit
on boot due to badly formatted /cpus nodes.
Downgrade this to a pr_warn() to be less alarmist, since it's not a
new problem.
Tested on Arndale, Cubox, Seaboard and Panda ES. Panda hits the WARN
without this, the others do not.
Acked-by: Russell King <rmk+kernel@arm.linux.org.uk>
Signed-off-by: Olof Johansson <olof@lixom.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit 37f02195b (powerpc/pci: fix PCI-e devices rescan issue on powerpc
platform) fixes a problem with interrupt and DMA initialization on hot
plugged devices. With this commit, interrupt and DMA initialization for
hot plugged devices is handled in the pci device enable function.
This approach has a couple of drawbacks. First, it creates two code paths
for device initialization, one for hot plugged devices and another for devices
known during the initial PCI scan. Second, the initialization code for hot
plugged devices is only called when the device is enabled, ie typically
in the probe function. Also, the platform specific setup code is called each
time pci_enable_device() is called, not only once during device discovery,
meaning it is actually called multiple times, once for devices discovered
during the initial scan and again each time a driver is re-loaded.
The visible result is that interrupt pins are only assigned to hot plugged
devices when the device driver is loaded. Effectively this changes the PCI
probe API, since pci_dev->irq and the device's dma configuration will now
only be valid after pci_enable() was called at least once. A more subtle
change is that platform specific PCI device setup is moved from device
discovery into the driver's probe function, more specifically into the
pci_enable_device() call.
To fix the inconsistencies, add new function pcibios_add_device.
Call pcibios_setup_device from pcibios_setup_bus_devices if device setup
is not complete, and from pcibios_add_device if bus setup is complete.
With this change, device setup code is moved back into device initialization,
and called exactly once for both static and hot plugged devices.
[ This also fixes a regression introduced by the above patch which
causes dev->irq to be overwritten under some cirumstances after
MSIs have been enabled for the device which leads to crashes due
to the MSI core "hijacking" dev->irq to store the base MSI number
and not the LSI. --BenH
]
Cc: Yuanquan Chen <Yuanquan.Chen@freescale.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Hiroo Matsumoto <matsumoto.hiroo@jp.fujitsu.com>
Signed-off-by: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Pull crypto fix from Herbert Xu:
"This fixes a crash in the crypto layer exposed by an SCTP test tool"
* git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6:
crypto: algboss - Hold ref count on larval
Pull drm/qxl fix from Dave Airlie:
"Bad me forgot an access check, possible security issue, but since this
is the first kernel with it, should be fine to just put it in now"
* 'drm-fixes' of git://people.freedesktop.org/~airlied/linux:
drm/qxl: add missing access check for execbuffer ioctl
Pull Ceph fix from Sage Weil:
"This is a recently spotted regression in the snapshot behavior...
It turns out several tests weren't being run in the nightlies so this
took a while to spot"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client:
rbd: send snapshot context with writes
Pull ubifs fixes from Al Viro:
"A couple of ubifs readdir/lseek race fixes. Stable fodder, really
nasty..."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
UBIFS: fix a horrid bug
UBIFS: prepare to fix a horrid bug
Pull two MN10300 fixes from David Howells:
"The first fixes a problem with passing arrays rather than pointers to
get_user() where __typeof__ then wants to declare and initialise an
array variable which gcc doesn't like.
The second fixes a problem whereby putting mem=xxx into the kernel
command line causes init=xxx to get an incorrect value."
* tag 'for-linus-20130628' of git://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-mn10300:
mn10300: Use early_param() to parse "mem=" parameter
mn10300: Allow to pass array name to get_user()
Pull timer fix from Thomas Gleixner:
"Correct an ordering issue in the tick broadcast code. I really wish
we'd get compensation for pain and suffering for each line of code we
write to work around dysfunctional timer hardware."
* 'timers-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
tick: Fix tick_broadcast_pending_mask not cleared
Pull perf fix from Ingo Molnar:
"One more fix for a recently discovered bug"
* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
perf: Disable monitoring on setuid processes for regular users
Al Viro pointed me to the fact that '->readdir()' and '->llseek()' have no
mutual exclusion, which means the 'ubifs_dir_llseek()' can be run while we are
in the middle of 'ubifs_readdir()'.
This means that 'file->private_data' can be freed while 'ubifs_readdir()' uses
it, and this is a very bad bug: not only 'ubifs_readdir()' can return garbage,
but this may corrupt memory and lead to all kinds of problems like crashes an
security holes.
This patch fixes the problem by using the 'file->f_version' field, which
'->llseek()' always unconditionally sets to zero. We set it to 1 in
'ubifs_readdir()' and whenever we detect that it became 0, we know there was a
seek and it is time to clear the state saved in 'file->private_data'.
I tested this patch by writing a user-space program which runds readdir and
seek in parallell. I could easily crash the kernel without these patches, but
could not crash it with these patches.
Cc: stable@vger.kernel.org
Reported-by: Al Viro <viro@zeniv.linux.org.uk>
Tested-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
