Trade one atomic op for a full memory barrier.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Get rid of the __call_single_node union and cleanup the API a little
to avoid external code relying on the structure layout as much.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Get rid of the __call_single_node union and clean up the API a little
to avoid external code relying on the structure layout as much.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
At fork time currently, a local node can be allowed to fill completely
and allow the periodic load balancer to fix the problem. This can be
problematic in cases where a task creates lots of threads that idle until
woken as part of a worker poll causing a memory bandwidth problem.
However, a "real" workload suffers badly from this behaviour. The workload
in question is mostly NUMA aware but spawns large numbers of threads
that act as a worker pool that can be called from anywhere. These need
to spread early to get reasonable behaviour.
This patch limits how much a local node can fill before spilling over
to another node and it will not be a universal win. Specifically,
very short-lived workloads that fit within a NUMA node would prefer
the memory bandwidth.
As I cannot describe the "real" workload, the best proxy measure I found
for illustration was a page fault microbenchmark. It's not representative
of the workload but demonstrates the hazard of the current behaviour.
pft timings
5.10.0-rc2 5.10.0-rc2
imbalancefloat-v2 forkspread-v2
Amean elapsed-1 46.37 ( 0.00%) 46.05 * 0.69%*
Amean elapsed-4 12.43 ( 0.00%) 12.49 * -0.47%*
Amean elapsed-7 7.61 ( 0.00%) 7.55 * 0.81%*
Amean elapsed-12 4.79 ( 0.00%) 4.80 ( -0.17%)
Amean elapsed-21 3.13 ( 0.00%) 2.89 * 7.74%*
Amean elapsed-30 3.65 ( 0.00%) 2.27 * 37.62%*
Amean elapsed-48 3.08 ( 0.00%) 2.13 * 30.69%*
Amean elapsed-79 2.00 ( 0.00%) 1.90 * 4.95%*
Amean elapsed-80 2.00 ( 0.00%) 1.90 * 4.70%*
This is showing the time to fault regions belonging to threads. The target
machine has 80 logical CPUs and two nodes. Note the ~30% gain when the
machine is approximately the point where one node becomes fully utilised.
The slower results are borderline noise.
Kernel building shows similar benefits around the same balance point.
Generally performance was either neutral or better in the tests conducted.
The main consideration with this patch is the point where fork stops
spreading a task so some workloads may benefit from different balance
points but it would be a risky tuning parameter.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20201120090630.3286-5-mgorman@techsingularity.net
Currently, an imbalance is only allowed when a destination node
is almost completely idle. This solved one basic class of problems
and was the cautious approach.
This patch revisits the possibility that NUMA nodes can be imbalanced
until 25% of the CPUs are occupied. The reasoning behind 25% is somewhat
superficial -- it's half the cores when HT is enabled. At higher
utilisations, balancing should continue as normal and keep things even
until scheduler domains are fully busy or over utilised.
Note that this is not expected to be a universal win. Any benchmark
that prefers spreading as wide as possible with limited communication
will favour the old behaviour as there is more memory bandwidth.
Workloads that communicate heavily in pairs such as netperf or tbench
benefit. For the tests I ran, the vast majority of workloads saw
a benefit so it seems to be a worthwhile trade-off.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20201120090630.3286-4-mgorman@techsingularity.net
In find_idlest_group(), the load imbalance is only relevant when the group
is either overloaded or fully busy but it is calculated unconditionally.
This patch moves the imbalance calculation to the context it is required.
Technically, it is a micro-optimisation but really the benefit is avoiding
confusing one type of imbalance with another depending on the group_type
in the next patch.
No functional change.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20201120090630.3286-3-mgorman@techsingularity.net
This is simply a preparation patch to make the following patches easier
to read. No functional change.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20201120090630.3286-2-mgorman@techsingularity.net
Now that the scheduler can deal with migrate disable properly, there is no
real compelling reason to make it only available for RT.
There are quite some code pathes which needlessly disable preemption in
order to prevent migration and some constructs like kmap_atomic() enforce
it implicitly.
Making it available independent of RT allows to provide a preemptible
variant of kmap_atomic() and makes the code more consistent in general.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Grudgingly-Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20201118204007.269943012@linutronix.de
In order to make accurate predictions across CPUs and for all performance
states, Energy Aware Scheduling (EAS) needs frequency-invariant load
tracking signals.
EAS task placement aims to minimize energy consumption, and does so in
part by limiting the search space to only CPUs with the highest spare
capacity (CPU capacity - CPU utilization) in their performance domain.
Those candidates are the placement choices that will keep frequency at
its lowest possible and therefore save the most energy.
But without frequency invariance, a CPU's utilization is relative to the
CPU's current performance level, and not relative to its maximum
performance level, which determines its capacity. As a result, it will
fail to correctly indicate any potential spare capacity obtained by an
increase in a CPU's performance level. Therefore, a non-invariant
utilization signal would render the EAS task placement logic invalid.
Now that we properly report support for the Frequency Invariance Engine
(FIE) through arch_scale_freq_invariant() for arm and arm64 systems,
while also ensuring a re-evaluation of the EAS use conditions for
possible invariance status change, we can assert this is the case when
initializing EAS. Warn and bail out otherwise.
Suggested-by: Quentin Perret <qperret@google.com>
Signed-off-by: Ionela Voinescu <ionela.voinescu@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20201027180713.7642-4-ionela.voinescu@arm.com
Task scheduler behavior depends on frequency invariance (FI) support and
the resulting invariant load tracking signals. For example, in order to
make accurate predictions across CPUs for all performance states, Energy
Aware Scheduling (EAS) needs frequency-invariant load tracking signals
and therefore it has a direct dependency on FI. This dependency is known,
but EAS enablement is not yet conditioned on the presence of FI during
the built of the scheduling domain hierarchy.
Before this is done, the following must be considered: while
arch_scale_freq_invariant() will see changes in FI support and could
be used to condition the use of EAS, it could return different values
during system initialisation.
For arm64, such a scenario will happen for a system that does not support
cpufreq driven FI, but does support counter-driven FI. For such a system,
arch_scale_freq_invariant() will return false if called before counter
based FI initialisation, but change its status to true after it.
If EAS becomes explicitly dependent on FI this would affect the task
scheduler behavior which builds its scheduling domain hierarchy well
before the late counter-based FI init. During that process, EAS would be
disabled due to its dependency on FI.
Two points of future early calls to arch_scale_freq_invariant() which
would determine EAS enablement are:
- (1) drivers/base/arch_topology.c:126 <<update_topology_flags_workfn>>
rebuild_sched_domains();
This will happen after CPU capacity initialisation.
- (2) kernel/sched/cpufreq_schedutil.c:917 <<rebuild_sd_workfn>>
rebuild_sched_domains_energy();
-->rebuild_sched_domains();
This will happen during sched_cpufreq_governor_change() for the
schedutil cpufreq governor.
Therefore, before enforcing the presence of FI support for the use of EAS,
ensure the following: if there is a change in FI support status after
counter init, use the existing rebuild_sched_domains_energy() function to
trigger a rebuild of the scheduling and performance domains that in turn
will determine the enablement of EAS.
Signed-off-by: Ionela Voinescu <ionela.voinescu@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Link: https://lkml.kernel.org/r/20201027180713.7642-3-ionela.voinescu@arm.com
Add the rebuild_sched_domains_energy() function to wrap the functionality
that rebuilds the scheduling domains if any of the Energy Aware Scheduling
(EAS) initialisation conditions change. This functionality is used when
schedutil is added or removed or when EAS is enabled or disabled
through the sched_energy_aware sysctl.
Therefore, create a single function that is used in both these cases and
that can be later reused.
Signed-off-by: Ionela Voinescu <ionela.voinescu@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Quentin Perret <qperret@google.com>
Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Link: https://lkml.kernel.org/r/20201027180713.7642-2-ionela.voinescu@arm.com
In case the user wants to stop controlling a uclamp constraint value
for a task, use the magic value -1 in sched_util_{min,max} with the
appropriate sched_flags (SCHED_FLAG_UTIL_CLAMP_{MIN,MAX}) to indicate
the reset.
The advantage over the 'additional flag' approach (i.e. introducing
SCHED_FLAG_UTIL_CLAMP_RESET) is that no additional flag has to be
exported via uapi. This avoids the need to document how this new flag
has be used in conjunction with the existing uclamp related flags.
The following subtle issue is fixed as well. When a uclamp constraint
value is set on a !user_defined uclamp_se it is currently first reset
and then set.
Fix this by AND'ing !user_defined with !SCHED_FLAG_UTIL_CLAMP which
stands for the 'sched class change' case.
The related condition 'if (uc_se->user_defined)' moved from
__setscheduler_uclamp() into uclamp_reset().
Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Yun Hsiang <hsiang023167@gmail.com>
Link: https://lkml.kernel.org/r/20201113113454.25868-1-dietmar.eggemann@arm.com
This document seems to be out of date for many, many years. Even it has
misspelled from the first day.
ARCH_HASH_SCHED_TUNE should be ARCH_HAS_SCHED_TUNE
ARCH_HASH_SCHED_DOMAIN should be ARCH_HAS_SCHED_DOMAIN
Since v2.6.14, kernel completely deleted the relevant code and even
arch_init_sched_domains() was deleted.
Right now, kernel is asking architectures to call set_sched_topology() to
override the default sched domains.
On the other hand, to print the schedule debug information, users need to
set sched_debug cmdline or enable it by sysfs entry. So this patch also
adds the description for sched_debug.
Signed-off-by: Barry Song <song.bao.hua@hisilicon.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Link: https://lkml.kernel.org/r/20201113115018.1628-1-song.bao.hua@hisilicon.com
NUMA topologies where the shortest path between some two nodes requires
three or more hops (i.e. diameter > 2) end up being misrepresented in the
scheduler topology structures.
This is currently detected when booting a kernel with CONFIG_SCHED_DEBUG=y
+ sched_debug on the cmdline, although this will only yield a warning about
sched_group spans not matching sched_domain spans:
ERROR: groups don't span domain->span
Add an explicit warning for that case, triggered regardless of
CONFIG_SCHED_DEBUG, and decorate it with an appropriate comment.
The topology described in the comment can be booted up on QEMU by appending
the following to your usual QEMU incantation:
-smp cores=4 \
-numa node,cpus=0,nodeid=0 -numa node,cpus=1,nodeid=1, \
-numa node,cpus=2,nodeid=2, -numa node,cpus=3,nodeid=3, \
-numa dist,src=0,dst=1,val=20, -numa dist,src=0,dst=2,val=30, \
-numa dist,src=0,dst=3,val=40, -numa dist,src=1,dst=2,val=20, \
-numa dist,src=1,dst=3,val=30, -numa dist,src=2,dst=3,val=20
A somewhat more realistic topology (6-node mesh) with the same affliction
can be conjured with:
-smp cores=6 \
-numa node,cpus=0,nodeid=0 -numa node,cpus=1,nodeid=1, \
-numa node,cpus=2,nodeid=2, -numa node,cpus=3,nodeid=3, \
-numa node,cpus=4,nodeid=4, -numa node,cpus=5,nodeid=5, \
-numa dist,src=0,dst=1,val=20, -numa dist,src=0,dst=2,val=30, \
-numa dist,src=0,dst=3,val=40, -numa dist,src=0,dst=4,val=30, \
-numa dist,src=0,dst=5,val=20, \
-numa dist,src=1,dst=2,val=20, -numa dist,src=1,dst=3,val=30, \
-numa dist,src=1,dst=4,val=20, -numa dist,src=1,dst=5,val=30, \
-numa dist,src=2,dst=3,val=20, -numa dist,src=2,dst=4,val=30, \
-numa dist,src=2,dst=5,val=40, \
-numa dist,src=3,dst=4,val=20, -numa dist,src=3,dst=5,val=30, \
-numa dist,src=4,dst=5,val=20
Signed-off-by: Valentin Schneider <valentin.schneider@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Link: https://lore.kernel.org/lkml/jhjtux5edo2.mognet@arm.com
One of our machines keeled over trying to rebuild the scheduler domains.
Mainline produces the same splat:
BUG: unable to handle page fault for address: 0000607f820054db
CPU: 2 PID: 149 Comm: kworker/1:1 Not tainted 5.10.0-rc1-master+ #6
Workqueue: events cpuset_hotplug_workfn
RIP: build_sched_domains
Call Trace:
partition_sched_domains_locked
rebuild_sched_domains_locked
cpuset_hotplug_workfn
It happens with cgroup2 and exclusive cpusets only. This reproducer
triggers it on an 8-cpu vm and works most effectively with no
preexisting child cgroups:
cd $UNIFIED_ROOT
mkdir cg1
echo 4-7 > cg1/cpuset.cpus
echo root > cg1/cpuset.cpus.partition
# with smt/control reading 'on',
echo off > /sys/devices/system/cpu/smt/control
RIP maps to
sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
from sd_init(). sd_id is calculated earlier in the same function:
cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
sd_id = cpumask_first(sched_domain_span(sd));
tl->mask(cpu), which reads cpu_sibling_map on x86, returns an empty mask
and so cpumask_first() returns >= nr_cpu_ids, which leads to the bogus
value from per_cpu_ptr() above.
The problem is a race between cpuset_hotplug_workfn() and a later
offline of CPU N. cpuset_hotplug_workfn() updates the effective masks
when N is still online, the offline clears N from cpu_sibling_map, and
then the worker uses the stale effective masks that still have N to
generate the scheduling domains, leading the worker to read
N's empty cpu_sibling_map in sd_init().
rebuild_sched_domains_locked() prevented the race during the cgroup2
cpuset series up until the Fixes commit changed its check. Make the
check more robust so that it can detect an offline CPU in any exclusive
cpuset's effective mask, not just the top one.
Fixes: 0ccea8feb9 ("cpuset: Make generate_sched_domains() work with partition")
Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Tejun Heo <tj@kernel.org>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/20201112171711.639541-1-daniel.m.jordan@oracle.com
Oleksandr reported hitting the WARN in the 'task_rq(p) != rq' branch
of migration_cpu_stop(). Valentin noted that using cpu_of(rq) in that
case is just plain wrong to begin with, since per the earlier branch
that isn't the actual CPU of the task.
Replace both instances of is_cpu_allowed() by a direct p->cpus_mask
test using task_cpu().
Reported-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Debugged-by: Valentin Schneider <valentin.schneider@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Qian reported that some fuzzer issuing sched_setaffinity() ends up stuck on
a wait_for_completion(). The problematic pattern seems to be:
affine_move_task()
// task_running() case
stop_one_cpu();
wait_for_completion(&pending->done);
Combined with, on the stopper side:
migration_cpu_stop()
// Task moved between unlocks and scheduling the stopper
task_rq(p) != rq &&
// task_running() case
dest_cpu >= 0
=> no complete_all()
This can happen with both PREEMPT and !PREEMPT, although !PREEMPT should
be more likely to see this given the targeted task has a much bigger window
to block and be woken up elsewhere before the stopper runs.
Make migration_cpu_stop() always look at pending affinity requests; signal
their completion if the stopper hits a rq mismatch but the task is
still within its allowed mask. When Migrate-Disable isn't involved, this
matches the previous set_cpus_allowed_ptr() vs migration_cpu_stop()
behaviour.
Fixes: 6d337eab04 ("sched: Fix migrate_disable() vs set_cpus_allowed_ptr()")
Reported-by: Qian Cai <cai@redhat.com>
Signed-off-by: Valentin Schneider <valentin.schneider@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/lkml/8b62fd1ad1b18def27f18e2ee2df3ff5b36d0762.camel@redhat.com
The CFS wakeup code will only ever go through EAS / its fast path on
"regular" wakeups (i.e. not on forks or execs). These are currently gated
by a check against 'sd_flag', which would be SD_BALANCE_WAKE at wakeup.
However, we now have a flag that explicitly tells us whether a wakeup is a
"regular" one, so hinge those conditions on that flag instead.
Signed-off-by: Valentin Schneider <valentin.schneider@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20201102184514.2733-4-valentin.schneider@arm.com
Only select_task_rq_fair() uses that parameter to do an actual domain
search, other classes only care about what kind of wakeup is happening
(fork, exec, or "regular") and thus just translate the flag into a wakeup
type.
WF_TTWU and WF_EXEC have just been added, use these along with WF_FORK to
encode the wakeup types we care about. For select_task_rq_fair(), we can
simply use the shiny new WF_flag : SD_flag mapping.
Signed-off-by: Valentin Schneider <valentin.schneider@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20201102184514.2733-3-valentin.schneider@arm.com
To remove the sd_flag parameter of select_task_rq(), we need another way of
encoding wakeup types. There already is a WF_FORK flag, add the missing two.
With that said, we still need an easy way to turn WF_foo into
SD_bar (e.g. WF_TTWU into SD_BALANCE_WAKE). As suggested by Peter, let's
make our lives easier and make them match exactly, and throw in some
compile-time checks for good measure.
Signed-off-by: Valentin Schneider <valentin.schneider@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20201102184514.2733-2-valentin.schneider@arm.com
Since ab93a4bc95 ("sched/fair: Remove distribute_running fromCFS
bandwidth"), there is nothing to protect between
raw_spin_lock_irqsave/store() in do_sched_cfs_slack_timer().
Signed-off-by: Hui Su <sh_def@163.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Phil Auld <pauld@redhat.com>
Reviewed-by: Ben Segall <bsegall@google.com>
Link: https://lkml.kernel.org/r/20201030144621.GA96974@rlk
In order to minimize the interference of migrate_disable() on lower
priority tasks, which can be deprived of runtime due to being stuck
below a higher priority task. Teach the RT/DL balancers to push away
these higher priority tasks when a lower priority task gets selected
to run on a freshly demoted CPU (pull).
This adds migration interference to the higher priority task, but
restores bandwidth to system that would otherwise be irrevocably lost.
Without this it would be possible to have all tasks on the system
stuck on a single CPU, each task preempted in a migrate_disable()
section with a single high priority task running.
This way we can still approximate running the M highest priority tasks
on the system.
Migrating the top task away is (ofcourse) still subject to
migrate_disable() too, which means the lower task is subject to an
interference equivalent to the worst case migrate_disable() section.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102347.499155098@infradead.org
There's a valid ->pi_lock recursion issue where the actual PI code
tries to wake up the stop task. Make lockdep aware so it doesn't
complain about this.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102347.406912197@infradead.org
We want migrate_disable() tasks to get PULLs in order for them to PUSH
away the higher priority task.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102347.310519774@infradead.org
Replace a bunch of cpumask_any*() instances with
cpumask_any*_distribute(), by injecting this little bit of random in
cpu selection, we reduce the chance two competing balance operations
working off the same lowest_mask pick the same CPU.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102347.190759694@infradead.org
On CPU unplug tasks which are in a migrate disabled region cannot be pushed
to a different CPU until they returned to migrateable state.
Account the number of tasks on a runqueue which are in a migrate disabled
section and make the hotplug wait mechanism respect that.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102347.067278757@infradead.org
Concurrent migrate_disable() and set_cpus_allowed_ptr() has
interesting features. We rely on set_cpus_allowed_ptr() to not return
until the task runs inside the provided mask. This expectation is
exported to userspace.
This means that any set_cpus_allowed_ptr() caller must wait until
migrate_enable() allows migrations.
At the same time, we don't want migrate_enable() to schedule, due to
patterns like:
preempt_disable();
migrate_disable();
...
migrate_enable();
preempt_enable();
And:
raw_spin_lock(&B);
spin_unlock(&A);
this means that when migrate_enable() must restore the affinity
mask, it cannot wait for completion thereof. Luck will have it that
that is exactly the case where there is a pending
set_cpus_allowed_ptr(), so let that provide storage for the async stop
machine.
Much thanks to Valentin who used TLA+ most effective and found lots of
'interesting' cases.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102346.921768277@infradead.org
Add the base migrate_disable() support (under protest).
While migrate_disable() is (currently) required for PREEMPT_RT, it is
also one of the biggest flaws in the system.
Notably this is just the base implementation, it is broken vs
sched_setaffinity() and hotplug, both solved in additional patches for
ease of review.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102346.818170844@infradead.org
Thread a u32 flags word through the *set_cpus_allowed*() callchain.
This will allow adding behavioural tweaks for future users.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102346.729082820@infradead.org
Since we now migrate tasks away before DYING, we should also move
bandwidth unthrottle, otherwise we can gain tasks from unthrottle
after we expect all tasks to be gone already.
Also; it looks like the RT balancers don't respect cpu_active() and
instead rely on rq->online in part, complete this. This too requires
we do set_rq_offline() earlier to match the cpu_active() semantics.
(The bigger patch is to convert RT to cpu_active() entirely)
Since set_rq_online() is called from sched_cpu_activate(), place
set_rq_offline() in sched_cpu_deactivate().
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102346.639538965@infradead.org
With the new mechanism which kicks tasks off the outgoing CPU at the end of
schedule() the situation on an outgoing CPU right before the stopper thread
brings it down completely is:
- All user tasks and all unbound kernel threads have either been migrated
away or are not running and the next wakeup will move them to a online CPU.
- All per CPU kernel threads, except cpu hotplug thread and the stopper
thread have either been unbound or parked by the responsible CPU hotplug
callback.
That means that at the last step before the stopper thread is invoked the
cpu hotplug thread is the last legitimate running task on the outgoing
CPU.
Add a final wait step right before the stopper thread is kicked which
ensures that any still running tasks on the way to park or on the way to
kick themself of the CPU are either sleeping or gone.
This allows to remove the migrate_tasks() crutch in sched_cpu_dying(). If
sched_cpu_dying() detects that there is still another running task aside of
the stopper thread then it will explode with the appropriate fireworks.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102346.547163969@infradead.org
Don't rely on the scheduler to force break affinity for us -- it will
stop doing that for per-cpu-kthreads.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102346.464718669@infradead.org
RT kernels need to ensure that all tasks which are not per CPU kthreads
have left the outgoing CPU to guarantee that no tasks are force migrated
within a migrate disabled section.
There is also some desire to (ab)use fine grained CPU hotplug control to
clear a CPU from active state to force migrate tasks which are not per CPU
kthreads away for power control purposes.
Add a mechanism which waits until all tasks which should leave the CPU
after the CPU active flag is cleared have moved to a different online CPU.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102346.377836842@infradead.org
In preparation for migrate_disable(), make sure only per-cpu kthreads
are allowed to run on !active CPUs.
This is ran (as one of the very first steps) from the cpu-hotplug
task which is a per-cpu kthread and completion of the hotplug
operation only requires such tasks.
This constraint enables the migrate_disable() implementation to wait
for completion of all migrate_disable regions on this CPU at hotplug
time without fear of any new ones starting.
This replaces the unlikely(rq->balance_callbacks) test at the tail of
context_switch with an unlikely(rq->balance_work), the fast path is
not affected.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102346.292709163@infradead.org
The intent of balance_callback() has always been to delay executing
balancing operations until the end of the current rq->lock section.
This is because balance operations must often drop rq->lock, and that
isn't safe in general.
However, as noted by Scott, there were a few holes in that scheme;
balance_callback() was called after rq->lock was dropped, which means
another CPU can interleave and touch the callback list.
Rework code to call the balance callbacks before dropping rq->lock
where possible, and otherwise splice the balance list onto a local
stack.
This guarantees that the balance list must be empty when we take
rq->lock. IOW, we'll only ever run our own balance callbacks.
Reported-by: Scott Wood <swood@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102346.203901269@infradead.org
Crashes in stop-machine are hard to connect to the calling code, add a
little something to help with that.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Link: https://lkml.kernel.org/r/20201023102346.116513635@infradead.org
As commit:
39f23ce07b ("sched/fair: Fix unthrottle_cfs_rq() for leaf_cfs_rq list")
does in unthrottle_cfs_rq(), throttle_cfs_rq() can also use the same
pattern as dequeue_task_fair().
No functional changes.
Signed-off-by: Peng Wang <rocking@linux.alibaba.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Phil Auld <pauld@redhat.com>
Cc: Ben Segall <bsegall@google.com>
Link: https://lore.kernel.org/r/f11dd2e3ab35cc538e2eb57bf0c99b6eaffce127.1604973978.git.rocking@linux.alibaba.com
In the case of a thread wakeup, wake_affine determines whether a core
will be chosen for the thread on the socket where the thread ran
previously or on the socket of the waker. This is done primarily by
comparing the load of the core where th thread ran previously (prev)
and the load of the waker (this).
commit 11f10e5420 ("sched/fair: Use load instead of runnable load
in wakeup path") changed the load computation from the runnable load
to the load average, where the latter includes the load of threads
that have already blocked on the core.
When a short-running daemon processes happens to run on prev, this
change raised the situation that prev could appear to have a greater
load than this, even when prev is actually idle. When prev and this
are on the same socket, the idle prev is detected later, in
select_idle_sibling. But if that does not hold, prev is completely
ignored, causing the waking thread to move to the socket of the waker.
In the case of N mostly active threads on N cores, this triggers other
migrations and hurts performance.
In contrast, before commit 11f10e5420, the load on an idle core
was 0, and in the case of a non-idle waker core, the effect of
wake_affine was to select prev as the target for searching for a core
for the waking thread.
To avoid unnecessary migrations, extend wake_affine_idle to check
whether the core where the thread previously ran is currently idle,
and if so simply return that core as the target.
[1] commit 11f10e5420 ("sched/fair: Use load instead of runnable
load in wakeup path")
This particularly has an impact when using the ondemand power manager,
where kworkers run every 0.004 seconds on all cores, increasing the
likelihood that an idle core will be considered to have a load.
The following numbers were obtained with the benchmarking tool
hyperfine (https://github.com/sharkdp/hyperfine) on the NAS parallel
benchmarks (https://www.nas.nasa.gov/publications/npb.html). The
tests were run on an 80-core Intel(R) Xeon(R) CPU E7-8870 v4 @
2.10GHz. Active (intel_pstate) and passive (intel_cpufreq) power
management were used. Times are in seconds. All experiments use all
160 hardware threads.
v5.9/intel-pstate v5.9+patch/intel-pstate
bt.C.c 24.725724+-0.962340 23.349608+-1.607214
lu.C.x 29.105952+-4.804203 25.249052+-5.561617
sp.C.x 31.220696+-1.831335 30.227760+-2.429792
ua.C.x 26.606118+-1.767384 25.778367+-1.263850
v5.9/ondemand v5.9+patch/ondemand
bt.C.c 25.330360+-1.028316 23.544036+-1.020189
lu.C.x 35.872659+-4.872090 23.719295+-3.883848
sp.C.x 32.141310+-2.289541 29.125363+-0.872300
ua.C.x 29.024597+-1.667049 25.728888+-1.539772
On the smaller data sets (A and B) and on the other NAS benchmarks
there is no impact on performance.
This also has a major impact on the splash2x.volrend benchmark of the
parsec benchmark suite that goes from 1m25 without this patch to 0m45,
in active (intel_pstate) mode.
Fixes: 11f10e5420 ("sched/fair: Use load instead of runnable load in wakeup path")
Signed-off-by: Julia Lawall <Julia.Lawall@inria.fr>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by Vincent Guittot <vincent.guittot@linaro.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Link: https://lkml.kernel.org/r/1603372550-14680-1-git-send-email-Julia.Lawall@inria.fr
Florian reported that all of kernel/sched/ is rebuild when
CONFIG_BLK_DEV_INITRD is changed, which, while not a bug is
unexpected. This is due to us including vmlinux.lds.h.
Jakub explained that the problem is that we put the alignment
requirement on the type instead of on a variable. Type alignment is a
minimum, the compiler is free to pick any larger alignment for a
specific instance of the type (eg. the variable).
So force the type alignment on all individual variable definitions and
remove the undesired dependency on vmlinux.lds.h.
Fixes: 85c2ce9104 ("sched, vmlinux.lds: Increase STRUCT_ALIGNMENT to 64 bytes for GCC-4.9")
Reported-by: Florian Fainelli <f.fainelli@gmail.com>
Suggested-by: Jakub Jelinek <jakub@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
do_sched_yield() invokes schedule() with interrupts disabled which is
not allowed. This goes back to the pre git era to commit a6efb709806c
("[PATCH] irqlock patch 2.5.27-H6") in the history tree.
Reenable interrupts and remove the misleading comment which "explains" it.
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/87r1pt7y5c.fsf@nanos.tec.linutronix.de
Document membarrier ordering scenarios in membarrier.c. Thanks to Alan
Stern for refreshing my memory. Now that I have those in mind, it seems
appropriate to serialize them to comments for posterity.
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20201020134715.13909-4-mathieu.desnoyers@efficios.com
Add comments and memory barrier to kthread_use_mm and kthread_unuse_mm
to allow the effect of membarrier(2) to apply to kthreads accessing
user-space memory as well.
Given that no prior kthread use this guarantee and that it only affects
kthreads, adding this guarantee does not affect user-space ABI.
Refine the check in membarrier_global_expedited to exclude runqueues
running the idle thread rather than all kthreads from the IPI cpumask.
Now that membarrier_global_expedited can IPI kthreads, the scheduler
also needs to update the runqueue's membarrier_state when entering lazy
TLB state.
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20201020134715.13909-3-mathieu.desnoyers@efficios.com
exit_mm should issue memory barriers after user-space memory accesses,
before clearing current->mm, to order user-space memory accesses
performed prior to exit_mm before clearing tsk->mm, which has the
effect of skipping the membarrier private expedited IPIs.
exit_mm should also update the runqueue's membarrier_state so
membarrier global expedited IPIs are not sent when they are not
needed.
The membarrier system call can be issued concurrently with do_exit
if we have thread groups created with CLONE_VM but not CLONE_THREAD.
Here is the scenario I have in mind:
Two thread groups are created, A and B. Thread group B is created by
issuing clone from group A with flag CLONE_VM set, but not CLONE_THREAD.
Let's assume we have a single thread within each thread group (Thread A
and Thread B).
The AFAIU we can have:
Userspace variables:
int x = 0, y = 0;
CPU 0 CPU 1
Thread A Thread B
(in thread group A) (in thread group B)
x = 1
barrier()
y = 1
exit()
exit_mm()
current->mm = NULL;
r1 = load y
membarrier()
skips CPU 0 (no IPI) because its current mm is NULL
r2 = load x
BUG_ON(r1 == 1 && r2 == 0)
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20201020134715.13909-2-mathieu.desnoyers@efficios.com
It is possible for find_new_ilb() to select the current CPU, however,
this only happens from newidle balancing, in which case need_resched()
will be true, and consequently nohz_csd_func() will not trigger the
softirq.
Exclude the current CPU from becoming an ILB target.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Add CPUPRI_HIGHER above the RT99 priority to denote the CPU is in use
by higher priority tasks (specifically deadline).
XXX: we should probably drive PUSH-PULL from cpupri, that would
automagically result in an RT-PUSH when DL sets cpupri to CPUPRI_HIGHER.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Peter Zijlstra