The throttle interaction made my brain hurt, make it consistently
about 0 transitions of h_nr_running.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
* Use simple CFS pick_task for DL pick_task
DL server's pick_task calls CFS's pick_next_task_fair(), this is wrong
because core scheduling's pick_task only calls CFS's pick_task() for
evaluation / checking of the CFS task (comparing across CPUs), not for
actually affirmatively picking the next task. This causes RB tree
corruption issues in CFS that were found by syzbot.
* Make pick_task_fair clear DL server
A DL task pick might set ->dl_server, but it is possible the task will
never run (say the other HT has a stop task). If the CFS task is picked
in the future directly (say without DL server), ->dl_server will be
set. So clear it in pick_task_fair().
This fixes the KASAN issue reported by syzbot in set_next_entity().
(DL refactoring suggestions by Vineeth Pillai).
Reported-by: Suleiman Souhlal <suleiman@google.com>
Signed-off-by: "Joel Fernandes (Google)" <joel@joelfernandes.org>
Signed-off-by: Daniel Bristot de Oliveira <bristot@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vineeth Pillai <vineeth@bitbyteword.org>
Tested-by: Juri Lelli <juri.lelli@redhat.com>
Link: https://lore.kernel.org/r/b10489ab1f03d23e08e6097acea47442e7d6466f.1716811044.git.bristot@kernel.org
Among the motivations for the DL servers is the real-time throttling
mechanism. This mechanism works by throttling the rt_rq after
running for a long period without leaving space for fair tasks.
The base dl server avoids this problem by boosting fair tasks instead
of throttling the rt_rq. The point is that it boosts without waiting
for potential starvation, causing some non-intuitive cases.
For example, an IRQ dispatches two tasks on an idle system, a fair
and an RT. The DL server will be activated, running the fair task
before the RT one. This problem can be avoided by deferring the
dl server activation.
By setting the defer option, the dl_server will dispatch an
SCHED_DEADLINE reservation with replenished runtime, but throttled.
The dl_timer will be set for the defer time at (period - runtime) ns
from start time. Thus boosting the fair rq at defer time.
If the fair scheduler has the opportunity to run while waiting
for defer time, the dl server runtime will be consumed. If
the runtime is completely consumed before the defer time, the
server will be replenished while still in a throttled state. Then,
the dl_timer will be reset to the new defer time
If the fair server reaches the defer time without consuming
its runtime, the server will start running, following CBS rules
(thus without breaking SCHED_DEADLINE). Then the server will
continue the running state (without deferring) until it fair
tasks are able to execute as regular fair scheduler (end of
the starvation).
Signed-off-by: Daniel Bristot de Oliveira <bristot@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Juri Lelli <juri.lelli@redhat.com>
Link: https://lore.kernel.org/r/dd175943c72533cd9f0b87767c6499204879cc38.1716811044.git.bristot@kernel.org
Consider the following cgroup:
root
|
------------------------
| |
normal_cgroup idle_cgroup
| |
SCHED_IDLE task_A SCHED_NORMAL task_B
According to the cgroup hierarchy, A should preempt B. But current
check_preempt_wakeup_fair() treats cgroup se and task separately, so B
will preempt A unexpectedly.
Unify the wakeup logic by {c,p}se_is_idle only. This makes SCHED_IDLE of
a task a relative policy that is effective only within its own cgroup,
similar to the behavior of NICE.
Also fix se_is_idle() definition when !CONFIG_FAIR_GROUP_SCHED.
Fixes: 304000390f ("sched: Cgroup SCHED_IDLE support")
Signed-off-by: Tianchen Ding <dtcccc@linux.alibaba.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Josh Don <joshdon@google.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20240626023505.1332596-1-dtcccc@linux.alibaba.com
When creating a new task, we initialize vruntime of the newly task at
sched_cgroup_fork(). However, the timing of executing this action is too
early and may not be accurate.
Because it uses current CPU to init the vruntime, but the new task
actually runs on the cpu which be assigned at wake_up_new_task().
To optimize this case, we pass ENQUEUE_INITIAL flag to activate_task()
in wake_up_new_task(), in this way, when place_entity is called in
enqueue_entity(), the vruntime of the new task will be initialized.
In addition, place_entity() in task_fork_fair() was introduced for two
reasons:
1. Previously, the __enqueue_entity() was in task_new_fair(),
in order to provide vruntime for enqueueing the newly task, the
vruntime assignment equation "se->vruntime = cfs_rq->min_vruntime" was
introduced by commit e9acbff648 ("sched: introduce se->vruntime").
This is the initial state of place_entity().
2. commit 4d78e7b656 ("sched: new task placement for vruntime") added
child_runs_first task placement feature which based on vruntime, this
also requires the new task's vruntime value.
After removing the child_runs_first and enqueue_entity() from
task_fork_fair(), this place_entity() no longer makes sense, so remove
it also.
Signed-off-by: Zhang Qiao <zhangqiao22@huawei.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20240627133359.1370598-1-zhangqiao22@huawei.com
d329605287 ("sched/fair: set_load_weight() must also call reweight_task()
for SCHED_IDLE tasks") applied to sched/core changes how reweight_task() is
called causing conflicts with e83edbf88f ("sched: Add
sched_class->reweight_task()"). Resolve the conflicts by taking
set_load_weight() changes from d329605287 and updating
sched_class->reweight_task() to take pointer to struct load_weight instead
of int prio.
Signed-off-by: Tejun Heo<tj@kernel.org>
When a task's weight is being changed, set_load_weight() is called with
@update_load set. As weight changes aren't trivial for the fair class,
set_load_weight() calls fair.c::reweight_task() for fair class tasks.
However, set_load_weight() first tests task_has_idle_policy() on entry and
skips calling reweight_task() for SCHED_IDLE tasks. This is buggy as
SCHED_IDLE tasks are just fair tasks with a very low weight and they would
incorrectly skip load, vlag and position updates.
Fix it by updating reweight_task() to take struct load_weight as idle weight
can't be expressed with prio and making set_load_weight() call
reweight_task() for SCHED_IDLE tasks too when @update_load is set.
Fixes: 9059393e4e ("sched/fair: Use reweight_entity() for set_user_nice()")
Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org # v4.15+
Link: http://lkml.kernel.org/r/20240624102331.GI31592@noisy.programming.kicks-ass.net
This reverts commit b0defa7ae0.
b0defa7ae0 changed the load balancing logic to ignore env.max_loop if
all tasks examined to that point were pinned. The goal of the patch was
to make it more likely to be able to detach a task buried in a long list
of pinned tasks. However, this has the unfortunate side effect of
creating an O(n) iteration in detach_tasks(), as we now must fully
iterate every task on a cpu if all or most are pinned. Since this load
balance code is done with rq lock held, and often in softirq context, it
is very easy to trigger hard lockups. We observed such hard lockups with
a user who affined O(10k) threads to a single cpu.
When I discussed this with Vincent he initially suggested that we keep
the limit on the number of tasks to detach, but increase the number of
tasks we can search. However, after some back and forth on the mailing
list, he recommended we instead revert the original patch, as it seems
likely no one was actually getting hit by the original issue.
Fixes: b0defa7ae0 ("sched/fair: Make sure to try to detach at least one movable task")
Signed-off-by: Josh Don <joshdon@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20240620214450.316280-1-joshdon@google.com
A new BPF extensible sched_class will need to dynamically change how a task
picks its sched_class. For example, if the loaded BPF scheduler progs fail,
the tasks will be forced back on CFS even if the task's policy is set to the
new sched_class. To support such mapping, add normal_policy() which wraps
testing for %SCHED_NORMAL. This doesn't cause any behavior changes.
v2: Update the description with more details on the expected use.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: David Vernet <dvernet@meta.com>
Acked-by: Josh Don <joshdon@google.com>
Acked-by: Hao Luo <haoluo@google.com>
Acked-by: Barret Rhoden <brho@google.com>
RT, DL, thermal and irq load and utilization metrics need to be decayed and
updated periodically and before consumption to keep the numbers reasonable.
This is currently done from __update_blocked_others() as a part of the fair
class load balance path. Let's factor it out to update_other_load_avgs().
Pure refactor. No functional changes.
This will be used by the new BPF extensible scheduling class to ensure that
the above metrics are properly maintained.
v2: Refreshed on top of tip:sched/core.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: David Vernet <dvernet@meta.com>
Currently, during a task weight change, sched core directly calls
reweight_task() defined in fair.c if @p is on CFS. Let's make it a proper
sched_class operation instead. CFS's reweight_task() is renamed to
reweight_task_fair() and now called through sched_class.
While it turns a direct call into an indirect one, set_load_weight() isn't
called from a hot path and this change shouldn't cause any noticeable
difference. This will be used to implement reweight_task for a new BPF
extensible sched_class so that it can keep its cached task weight
up-to-date.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: David Vernet <dvernet@meta.com>
Acked-by: Josh Don <joshdon@google.com>
Acked-by: Hao Luo <haoluo@google.com>
Acked-by: Barret Rhoden <brho@google.com>
We observed that the overhead on trigger_load_balance(), now renamed
sched_balance_trigger(), has risen with a system's core counts.
For an OLTP workload running 6.8 kernel on a 2 socket x86 systems
having 96 cores/socket, we saw that 0.7% cpu cycles are spent in
trigger_load_balance(). On older systems with fewer cores/socket, this
function's overhead was less than 0.1%.
The cause of this overhead was that there are multiple cpus calling
kick_ilb(flags), updating the balancing work needed to a common idle
load balancer cpu. The ilb_cpu's flags field got updated unconditionally
with atomic_fetch_or(). The atomic read and writes to ilb_cpu's flags
causes much cache bouncing and cpu cycles overhead. This is seen in the
annotated profile below.
kick_ilb():
if (ilb_cpu < 0)
test %r14d,%r14d
↑ js 6c
flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
mov $0x2d600,%rdi
movslq %r14d,%r8
mov %rdi,%rdx
add -0x7dd0c3e0(,%r8,8),%rdx
arch_atomic_read():
0.01 mov 0x64(%rdx),%esi
35.58 add $0x64,%rdx
arch_atomic_fetch_or():
static __always_inline int arch_atomic_fetch_or(int i, atomic_t *v)
{
int val = arch_atomic_read(v);
do { } while (!arch_atomic_try_cmpxchg(v, &val, val | i));
0.03 157: mov %r12d,%ecx
arch_atomic_try_cmpxchg():
return arch_try_cmpxchg(&v->counter, old, new);
0.00 mov %esi,%eax
arch_atomic_fetch_or():
do { } while (!arch_atomic_try_cmpxchg(v, &val, val | i));
or %esi,%ecx
arch_atomic_try_cmpxchg():
return arch_try_cmpxchg(&v->counter, old, new);
0.01 lock cmpxchg %ecx,(%rdx)
42.96 ↓ jne 2d2
kick_ilb():
With instrumentation, we found that 81% of the updates do not result in
any change in the ilb_cpu's flags. That is, multiple cpus are asking
the ilb_cpu to do the same things over and over again, before the ilb_cpu
has a chance to run NOHZ load balance.
Skip updates to ilb_cpu's flags if no new work needs to be done.
Such updates do not change ilb_cpu's NOHZ flags. This requires an extra
atomic read but it is less expensive than frequent unnecessary atomic
updates that generate cache bounces.
We saw that on the OLTP workload, cpu cycles from trigger_load_balance()
(or sched_balance_trigger()) got reduced from 0.7% to 0.2%.
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Chen Yu <yu.c.chen@intel.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20240531205452.65781-1-tim.c.chen@linux.intel.com
Pull sysctl updates from Joel Granados:
- Remove sentinel elements from ctl_table structs in kernel/*
Removing sentinels in ctl_table arrays reduces the build time size
and runtime memory consumed by ~64 bytes per array. Removals for
net/, io_uring/, mm/, ipc/ and security/ are set to go into mainline
through their respective subsystems making the next release the most
likely place where the final series that removes the check for
proc_name == NULL will land.
This adds to removals already in arch/, drivers/ and fs/.
- Adjust ctl_table definitions and references to allow constification
- Remove unused ctl_table function arguments
- Move non-const elements from ctl_table to ctl_table_header
- Make ctl_table pointers const in ctl_table_root structure
Making the static ctl_table structs const will increase safety by
keeping the pointers to proc_handler functions in .rodata. Though no
ctl_tables where made const in this PR, the ground work for making
that possible has started with these changes sent by Thomas
Weißschuh.
* tag 'sysctl-6.10-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/sysctl/sysctl:
sysctl: drop now unnecessary out-of-bounds check
sysctl: move sysctl type to ctl_table_header
sysctl: drop sysctl_is_perm_empty_ctl_table
sysctl: treewide: constify argument ctl_table_root::permissions(table)
sysctl: treewide: drop unused argument ctl_table_root::set_ownership(table)
bpf: Remove the now superfluous sentinel elements from ctl_table array
delayacct: Remove the now superfluous sentinel elements from ctl_table array
kprobes: Remove the now superfluous sentinel elements from ctl_table array
printk: Remove the now superfluous sentinel elements from ctl_table array
scheduler: Remove the now superfluous sentinel elements from ctl_table array
seccomp: Remove the now superfluous sentinel elements from ctl_table array
timekeeping: Remove the now superfluous sentinel elements from ctl_table array
ftrace: Remove the now superfluous sentinel elements from ctl_table array
umh: Remove the now superfluous sentinel elements from ctl_table array
kernel misc: Remove the now superfluous sentinel elements from ctl_table array
On 05/03/2024 15:05, Vincent Guittot wrote:
I'm fine with either and that was my first thought here, too, but it did seem like
the comment was mostly placed there to justify the 'unexpected' high utilization
when explicitly passing FREQUENCY_UTIL and the need to clamp it then.
So removing did feel slightly more natural to me anyway.
So alternatively:
From: Christian Loehle <christian.loehle@arm.com>
Date: Tue, 5 Mar 2024 09:34:41 +0000
Subject: [PATCH] sched/fair: Remove stale FREQUENCY_UTIL mention
effective_cpu_util() flags were removed, so remove mentioning of the
flag.
commit 9c0b4bb7f6 ("sched/cpufreq: Rework schedutil governor performance estimation")
reworked effective_cpu_util() removing enum cpu_util_type. Modify the
comment accordingly.
Signed-off-by: Christian Loehle <christian.loehle@arm.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/0e2833ee-0939-44e0-82a2-520a585a0153@arm.com
Change se->load.weight to se_weight(se) in the calculation for the
initial util_avg to avoid unnecessarily inflating the util_avg by 1024
times.
The reason is that se->load.weight has the unit/scale as the scaled-up
load, while cfs_rg->avg.load_avg has the unit/scale as the true task
weight (as mapped directly from the task's nice/priority value). With
CONFIG_32BIT, the scaled-up load is equal to the true task weight. With
CONFIG_64BIT, the scaled-up load is 1024 times the true task weight.
Thus, the current code may inflate the util_avg by 1024 times. The
follow-up capping will not allow the util_avg value to go wild. But the
calculation should have the correct logic.
Signed-off-by: Dawei Li <daweilics@gmail.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Vishal Chourasia <vishalc@linux.ibm.com>
Link: https://lore.kernel.org/r/20240315015916.21545-1-daweilics@gmail.com
Pull scheduler updates from Ingo Molnar:
- Add cpufreq pressure feedback for the scheduler
- Rework misfit load-balancing wrt affinity restrictions
- Clean up and simplify the code around ::overutilized and
::overload access.
- Simplify sched_balance_newidle()
- Bump SCHEDSTAT_VERSION to 16 due to a cleanup of CPU_MAX_IDLE_TYPES
handling that changed the output.
- Rework & clean up <asm/vtime.h> interactions wrt arch_vtime_task_switch()
- Reorganize, clean up and unify most of the higher level
scheduler balancing function names around the sched_balance_*()
prefix
- Simplify the balancing flag code (sched_balance_running)
- Miscellaneous cleanups & fixes
* tag 'sched-core-2024-05-13' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (50 commits)
sched/pelt: Remove shift of thermal clock
sched/cpufreq: Rename arch_update_thermal_pressure() => arch_update_hw_pressure()
thermal/cpufreq: Remove arch_update_thermal_pressure()
sched/cpufreq: Take cpufreq feedback into account
cpufreq: Add a cpufreq pressure feedback for the scheduler
sched/fair: Fix update of rd->sg_overutilized
sched/vtime: Do not include <asm/vtime.h> header
s390/irq,nmi: Include <asm/vtime.h> header directly
s390/vtime: Remove unused __ARCH_HAS_VTIME_TASK_SWITCH leftover
sched/vtime: Get rid of generic vtime_task_switch() implementation
sched/vtime: Remove confusing arch_vtime_task_switch() declaration
sched/balancing: Simplify the sg_status bitmask and use separate ->overloaded and ->overutilized flags
sched/fair: Rename set_rd_overutilized_status() to set_rd_overutilized()
sched/fair: Rename SG_OVERLOAD to SG_OVERLOADED
sched/fair: Rename {set|get}_rd_overload() to {set|get}_rd_overloaded()
sched/fair: Rename root_domain::overload to ::overloaded
sched/fair: Use helper functions to access root_domain::overload
sched/fair: Check root_domain::overload value before update
sched/fair: Combine EAS check with root_domain::overutilized access
sched/fair: Simplify the continue_balancing logic in sched_balance_newidle()
...
Now that cpufreq provides a pressure value to the scheduler, rename
arch_update_thermal_pressure into HW pressure to reflect that it returns
a pressure applied by HW (i.e. with a high frequency change) and not
always related to thermal mitigation but also generated by max current
limitation as an example. Such high frequency signal needs filtering to be
smoothed and provide an value that reflects the average available capacity
into the scheduler time scale.
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Tested-by: Lukasz Luba <lukasz.luba@arm.com>
Reviewed-by: Qais Yousef <qyousef@layalina.io>
Reviewed-by: Lukasz Luba <lukasz.luba@arm.com>
Link: https://lore.kernel.org/r/20240326091616.3696851-5-vincent.guittot@linaro.org
SG_OVERLOADED and SG_OVERUTILIZED flags plus the sg_status bitmask are an
unnecessary complication that only make the code harder to read and slower.
We only ever set them separately:
thule:~/tip> git grep SG_OVER kernel/sched/
kernel/sched/fair.c: set_rd_overutilized_status(rq->rd, SG_OVERUTILIZED);
kernel/sched/fair.c: *sg_status |= SG_OVERLOADED;
kernel/sched/fair.c: *sg_status |= SG_OVERUTILIZED;
kernel/sched/fair.c: *sg_status |= SG_OVERLOADED;
kernel/sched/fair.c: set_rd_overloaded(env->dst_rq->rd, sg_status & SG_OVERLOADED);
kernel/sched/fair.c: sg_status & SG_OVERUTILIZED);
kernel/sched/fair.c: } else if (sg_status & SG_OVERUTILIZED) {
kernel/sched/fair.c: set_rd_overutilized_status(env->dst_rq->rd, SG_OVERUTILIZED);
kernel/sched/sched.h:#define SG_OVERLOADED 0x1 /* More than one runnable task on a CPU. */
kernel/sched/sched.h:#define SG_OVERUTILIZED 0x2 /* One or more CPUs are over-utilized. */
kernel/sched/sched.h: set_rd_overloaded(rq->rd, SG_OVERLOADED);
And use them separately, which results in suboptimal code:
/* update overload indicator if we are at root domain */
set_rd_overloaded(env->dst_rq->rd, sg_status & SG_OVERLOADED);
/* Update over-utilization (tipping point, U >= 0) indicator */
set_rd_overutilized_status(env->dst_rq->rd,
Introduce separate sg_overloaded and sg_overutilized flags in update_sd_lb_stats()
and its lower level functions, and change all of them to 'bool'.
Remove the now unused SG_OVERLOADED and SG_OVERUTILIZED flags.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Acked-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Tested-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Cc: Qais Yousef <qyousef@layalina.io>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/ZgVPhODZ8/nbsqbP@gmail.com
The root_domain::overload flag is 1 when there's any rq
in the root domain that has 2 or more running tasks. (Ie. it's overloaded.)
The root_domain structure itself is a global structure per cpuset island.
The ::overload flag is maintained the following way:
- Set when adding a second task to the runqueue.
- It is cleared in update_sd_lb_stats() during load balance,
if none of the rqs have 2 or more running tasks.
This flag is used during newidle balance to see if its worth doing a full
load balance pass, which can be an expensive operation. If it is set,
then newidle balance will try to aggressively pull a task.
Since commit:
630246a06a ("sched/fair: Clean-up update_sg_lb_stats parameters")
::overload is being written unconditionally, even if it has the same
value. The change in value of this depends on the workload, but on
typical workloads, it doesn't change all that often: a system is
either dominantly overloaded for substantial amounts of time, or not.
Extra writes to this semi-global structure cause unnecessary overhead, extra
bus traffic, etc. - so avoid it as much as possible.
Perf probe stats show that it's worth making this change (numbers are
with patch applied):
1M probe:sched_balance_newidle_L38
139 probe:update_sd_lb_stats_L53 <====== 1->0 writes
129K probe:add_nr_running_L12
74 probe:add_nr_running_L13 <====== 0->1 writes
54K probe:update_sd_lb_stats_L50 <====== reads
These numbers prove that actual change in the ::overload value is (much) less
frequent: L50 is much larger at ~54,000 accesses vs L53+L13 of 139+74.
[ mingo: Rewrote the changelog. ]
Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Qais Yousef <qyousef@layalina.io>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Link: https://lore.kernel.org/r/20240325054505.201995-2-sshegde@linux.ibm.com
Access to root_domainoverutilized is always used with sched_energy_enabled in
the pattern:
if (sched_energy_enabled && !overutilized)
do something
So modify the helper function to utilize this pattern. This is more
readable code as it would say, do something when root domain is not
overutilized. This function always return true when EAS is disabled.
No change in functionality intended.
Suggested-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20240326152616.380999-1-sshegde@linux.ibm.com
newidle(CPU_NEWLY_IDLE) balancing doesn't stop the load-balancing if the
continue_balancing flag is reset, but the other two balancing (IDLE, BUSY)
cases do that.
newidle balance stops the load balancing if rq has a task or there
is wakeup pending. The same checks are present in should_we_balance for
newidle. Hence use the return value and simplify continue_balancing
mechanism for newidle. Update the comment surrounding it as well.
No change in functionality intended.
Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Link: https://lore.kernel.org/r/20240325153926.274284-1-sshegde@linux.ibm.com
root_domain::overutilized is only used for EAS(energy aware scheduler)
to decide whether to do load balance or not. It is not used if EAS
not possible.
Currently enqueue_task_fair and task_tick_fair accesses, sometime updates
this field. In update_sd_lb_stats it is updated often. This causes cache
contention due to true sharing and burns a lot of cycles. ::overload and
::overutilized are part of the same cacheline. Updating it often invalidates
the cacheline. That causes access to ::overload to slow down due to
false sharing. Hence add EAS check before accessing/updating this field.
EAS check is optimized at compile time or it is a static branch.
Hence it shouldn't cost much.
With the patch, both enqueue_task_fair and newidle_balance don't show
up as hot routines in perf profile.
6.8-rc4:
7.18% swapper [kernel.vmlinux] [k] enqueue_task_fair
6.78% s [kernel.vmlinux] [k] newidle_balance
+patch:
0.14% swapper [kernel.vmlinux] [k] enqueue_task_fair
0.00% swapper [kernel.vmlinux] [k] newidle_balance
While at it: trace_sched_overutilized_tp expect that second argument to
be bool. So do a int to bool conversion for that.
Fixes: 2802bf3cd9 ("sched/fair: Add over-utilization/tipping point indicator")
Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Qais Yousef <qyousef@layalina.io>
Reviewed-by: Srikar Dronamraju <srikar@linux.ibm.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20240307085725.444486-2-sshegde@linux.ibm.com
It is not necessarily an indication of the system being busy and
requires a backoff of the load balancer activities. But pushing it high
could mean generally delaying other misfit activities or other type of
imbalances.
Also don't pollute nr_balance_failed because of misfit failures. The
value is used for enabling cache hot migration and in migrate_util/load
types. None of which should be impacted (skewed) by misfit failures.
Signed-off-by: Qais Yousef <qyousef@layalina.io>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20240324004552.999936-5-qyousef@layalina.io
If a misfit task is affined to a subset of the possible CPUs, we need to
verify that one of these CPUs can fit it. Otherwise the load balancer
code will continuously trigger needlessly leading the balance_interval
to increase in return and eventually end up with a situation where real
imbalances take a long time to address because of this impossible
imbalance situation.
This can happen in Android world where it's common for background tasks
to be restricted to little cores.
Similarly if we can't fit the biggest core, triggering misfit is
pointless as it is the best we can ever get on this system.
To be able to detect that; we use asym_cap_list to iterate through
capacities in the system to see if the task is able to run at a higher
capacity level based on its p->cpus_ptr. We do that when the affinity
change, a fair task is forked, or when a task switched to fair policy.
We store the max_allowed_capacity in task_struct to allow for cheap
comparison in the fast path.
Improve check_misfit_status() function by removing redundant checks.
misfit_task_load will be 0 if the task can't move to a bigger CPU. And
nohz_balancer_kick() already checks for cpu_check_capacity() before
calling check_misfit_status().
Test:
=====
Add
trace_printk("balance_interval = %lu\n", interval)
in get_sd_balance_interval().
run
if [ "$MASK" != "0" ]; then
adb shell "taskset -a $MASK cat /dev/zero > /dev/null"
fi
sleep 10
// parse ftrace buffer counting the occurrence of each valaue
Where MASK is either:
* 0: no busy task running
* 1: busy task is pinned to 1 cpu; handled today to not cause
misfit
* f: busy task pinned to little cores, simulates busy background
task, demonstrates the problem to be fixed
Results:
========
Note how occurrence of balance_interval = 128 overshoots for MASK = f.
BEFORE
------
MASK=0
1 balance_interval = 175
120 balance_interval = 128
846 balance_interval = 64
55 balance_interval = 63
215 balance_interval = 32
2 balance_interval = 31
2 balance_interval = 16
4 balance_interval = 8
1870 balance_interval = 4
65 balance_interval = 2
MASK=1
27 balance_interval = 175
37 balance_interval = 127
840 balance_interval = 64
167 balance_interval = 63
449 balance_interval = 32
84 balance_interval = 31
304 balance_interval = 16
1156 balance_interval = 8
2781 balance_interval = 4
428 balance_interval = 2
MASK=f
1 balance_interval = 175
1328 balance_interval = 128
44 balance_interval = 64
101 balance_interval = 63
25 balance_interval = 32
5 balance_interval = 31
23 balance_interval = 16
23 balance_interval = 8
4306 balance_interval = 4
177 balance_interval = 2
AFTER
-----
Note how the high values almost disappear for all MASK values. The
system has background tasks that could trigger the problem without
simulate it even with MASK=0.
MASK=0
103 balance_interval = 63
19 balance_interval = 31
194 balance_interval = 8
4827 balance_interval = 4
179 balance_interval = 2
MASK=1
131 balance_interval = 63
1 balance_interval = 31
87 balance_interval = 8
3600 balance_interval = 4
7 balance_interval = 2
MASK=f
8 balance_interval = 127
182 balance_interval = 63
3 balance_interval = 31
9 balance_interval = 16
415 balance_interval = 8
3415 balance_interval = 4
21 balance_interval = 2
Signed-off-by: Qais Yousef <qyousef@layalina.io>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20240324004552.999936-3-qyousef@layalina.io
Pull MM updates from Andrew Morton:
- Sumanth Korikkar has taught s390 to allocate hotplug-time page frames
from hotplugged memory rather than only from main memory. Series
"implement "memmap on memory" feature on s390".
- More folio conversions from Matthew Wilcox in the series
"Convert memcontrol charge moving to use folios"
"mm: convert mm counter to take a folio"
- Chengming Zhou has optimized zswap's rbtree locking, providing
significant reductions in system time and modest but measurable
reductions in overall runtimes. The series is "mm/zswap: optimize the
scalability of zswap rb-tree".
- Chengming Zhou has also provided the series "mm/zswap: optimize zswap
lru list" which provides measurable runtime benefits in some
swap-intensive situations.
- And Chengming Zhou further optimizes zswap in the series "mm/zswap:
optimize for dynamic zswap_pools". Measured improvements are modest.
- zswap cleanups and simplifications from Yosry Ahmed in the series
"mm: zswap: simplify zswap_swapoff()".
- In the series "Add DAX ABI for memmap_on_memory", Vishal Verma has
contributed several DAX cleanups as well as adding a sysfs tunable to
control the memmap_on_memory setting when the dax device is
hotplugged as system memory.
- Johannes Weiner has added the large series "mm: zswap: cleanups",
which does that.
- More DAMON work from SeongJae Park in the series
"mm/damon: make DAMON debugfs interface deprecation unignorable"
"selftests/damon: add more tests for core functionalities and corner cases"
"Docs/mm/damon: misc readability improvements"
"mm/damon: let DAMOS feeds and tame/auto-tune itself"
- In the series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension" Rakie Kim has developed a new mempolicy interleaving
policy wherein we allocate memory across nodes in a weighted fashion
rather than uniformly. This is beneficial in heterogeneous memory
environments appearing with CXL.
- Christophe Leroy has contributed some cleanup and consolidation work
against the ARM pagetable dumping code in the series "mm: ptdump:
Refactor CONFIG_DEBUG_WX and check_wx_pages debugfs attribute".
- Luis Chamberlain has added some additional xarray selftesting in the
series "test_xarray: advanced API multi-index tests".
- Muhammad Usama Anjum has reworked the selftest code to make its
human-readable output conform to the TAP ("Test Anything Protocol")
format. Amongst other things, this opens up the use of third-party
tools to parse and process out selftesting results.
- Ryan Roberts has added fork()-time PTE batching of THP ptes in the
series "mm/memory: optimize fork() with PTE-mapped THP". Mainly
targeted at arm64, this significantly speeds up fork() when the
process has a large number of pte-mapped folios.
- David Hildenbrand also gets in on the THP pte batching game in his
series "mm/memory: optimize unmap/zap with PTE-mapped THP". It
implements batching during munmap() and other pte teardown
situations. The microbenchmark improvements are nice.
- And in the series "Transparent Contiguous PTEs for User Mappings"
Ryan Roberts further utilizes arm's pte's contiguous bit ("contpte
mappings"). Kernel build times on arm64 improved nicely. Ryan's
series "Address some contpte nits" provides some followup work.
- In the series "mm/hugetlb: Restore the reservation" Breno Leitao has
fixed an obscure hugetlb race which was causing unnecessary page
faults. He has also added a reproducer under the selftest code.
- In the series "selftests/mm: Output cleanups for the compaction
test", Mark Brown did what the title claims.
- Kinsey Ho has added the series "mm/mglru: code cleanup and
refactoring".
- Even more zswap material from Nhat Pham. The series "fix and extend
zswap kselftests" does as claimed.
- In the series "Introduce cpu_dcache_is_aliasing() to fix DAX
regression" Mathieu Desnoyers has cleaned up and fixed rather a mess
in our handling of DAX on archiecctures which have virtually aliasing
data caches. The arm architecture is the main beneficiary.
- Lokesh Gidra's series "per-vma locks in userfaultfd" provides
dramatic improvements in worst-case mmap_lock hold times during
certain userfaultfd operations.
- Some page_owner enhancements and maintenance work from Oscar Salvador
in his series
"page_owner: print stacks and their outstanding allocations"
"page_owner: Fixup and cleanup"
- Uladzislau Rezki has contributed some vmalloc scalability
improvements in his series "Mitigate a vmap lock contention". It
realizes a 12x improvement for a certain microbenchmark.
- Some kexec/crash cleanup work from Baoquan He in the series "Split
crash out from kexec and clean up related config items".
- Some zsmalloc maintenance work from Chengming Zhou in the series
"mm/zsmalloc: fix and optimize objects/page migration"
"mm/zsmalloc: some cleanup for get/set_zspage_mapping()"
- Zi Yan has taught the MM to perform compaction on folios larger than
order=0. This a step along the path to implementaton of the merging
of large anonymous folios. The series is named "Enable >0 order folio
memory compaction".
- Christoph Hellwig has done quite a lot of cleanup work in the
pagecache writeback code in his series "convert write_cache_pages()
to an iterator".
- Some modest hugetlb cleanups and speedups in Vishal Moola's series
"Handle hugetlb faults under the VMA lock".
- Zi Yan has changed the page splitting code so we can split huge pages
into sizes other than order-0 to better utilize large folios. The
series is named "Split a folio to any lower order folios".
- David Hildenbrand has contributed the series "mm: remove
total_mapcount()", a cleanup.
- Matthew Wilcox has sought to improve the performance of bulk memory
freeing in his series "Rearrange batched folio freeing".
- Gang Li's series "hugetlb: parallelize hugetlb page init on boot"
provides large improvements in bootup times on large machines which
are configured to use large numbers of hugetlb pages.
- Matthew Wilcox's series "PageFlags cleanups" does that.
- Qi Zheng's series "minor fixes and supplement for ptdesc" does that
also. S390 is affected.
- Cleanups to our pagemap utility functions from Peter Xu in his series
"mm/treewide: Replace pXd_large() with pXd_leaf()".
- Nico Pache has fixed a few things with our hugepage selftests in his
series "selftests/mm: Improve Hugepage Test Handling in MM
Selftests".
- Also, of course, many singleton patches to many things. Please see
the individual changelogs for details.
* tag 'mm-stable-2024-03-13-20-04' of git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm: (435 commits)
mm/zswap: remove the memcpy if acomp is not sleepable
crypto: introduce: acomp_is_async to expose if comp drivers might sleep
memtest: use {READ,WRITE}_ONCE in memory scanning
mm: prohibit the last subpage from reusing the entire large folio
mm: recover pud_leaf() definitions in nopmd case
selftests/mm: skip the hugetlb-madvise tests on unmet hugepage requirements
selftests/mm: skip uffd hugetlb tests with insufficient hugepages
selftests/mm: dont fail testsuite due to a lack of hugepages
mm/huge_memory: skip invalid debugfs new_order input for folio split
mm/huge_memory: check new folio order when split a folio
mm, vmscan: retry kswapd's priority loop with cache_trim_mode off on failure
mm: add an explicit smp_wmb() to UFFDIO_CONTINUE
mm: fix list corruption in put_pages_list
mm: remove folio from deferred split list before uncharging it
filemap: avoid unnecessary major faults in filemap_fault()
mm,page_owner: drop unnecessary check
mm,page_owner: check for null stack_record before bumping its refcount
mm: swap: fix race between free_swap_and_cache() and swapoff()
mm/treewide: align up pXd_leaf() retval across archs
mm/treewide: drop pXd_large()
...
Standardize scheduler load-balancing function names on the
sched_balance_() prefix.
Also use 'dst' instead of 'idlest', because it's not really
true that we return the 'idlest' group or CPU, we sort by
idle-exit latency and only return the idlest CPUs from the
lowest-latency set of CPUs.
The true 'idlest' CPUs often remain idle for a long time
and are never returned as long as the system is under-loaded.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Link: https://lore.kernel.org/r/20240308111819.1101550-14-mingo@kernel.org
Standardize scheduler load-balancing function names on the
sched_balance_() prefix.
Also use 'dst' instead of 'idlest', because it's not really
true that we return the 'idlest' group or CPU, we sort by
idle-exit latency and only return the idlest CPUs from the
lowest-latency set of CPUs.
The true 'idlest' CPUs often remain idle for a long time
and are never returned as long as the system is under-loaded.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Link: https://lore.kernel.org/r/20240308111819.1101550-13-mingo@kernel.org
Standardize scheduler load-balancing function names on the
sched_balance_() prefix.
Also use 'dst' instead of 'idlest': while historically correct,
today it's not really true anymore that we return the 'idlest'
group or CPU, we sort by idle-exit latency and only return the
idlest CPUs from the lowest-latency set of CPUs.
The true 'idlest' CPUs often remain idle for a long time
and are never returned as long as the system is under-loaded.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Link: https://lore.kernel.org/r/20240308111819.1101550-12-mingo@kernel.org
