- 07 Oct, 2021 4 commits
-
-
Peter Zijlstra authored
Simplify and make wake_up_if_idle() more robust, also don't iterate the whole machine with preempt_disable() in it's caller: wake_up_all_idle_cpus(). This prepares for another wake_up_if_idle() user that needs a full do_idle() cycle. Signed-off-by:
Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by:
Vasily Gorbik <gor@linux.ibm.com> Tested-by: Vasily Gorbik <gor@linux.ibm.com> # on s390 Link: https://lkml.kernel.org/r/20210929152428.769328779@infradead.org
-
Peter Zijlstra authored
Instead of frobbing around with scheduler internals, use the shiny new task_call_func() interface. Signed-off-by:
Petr Mladek <pmladek@suse.com> Acked-by:
Miroslav Benes <mbenes@suse.cz> Acked-by:
-
Peter Zijlstra authored
Give try_invoke_on_locked_down_task() a saner name and have it return an int so that the caller might distinguish between different reasons of failure. Signed-off-by:
Paul E. McKenney <paulmck@kernel.org> Acked-by:
-
Peter Zijlstra authored
Clarify and tighten try_invoke_on_locked_down_task(). Basically the function calls @func under task_rq_lock(), except it avoids taking rq->lock when possible. This makes calling @func unconditional (the function will get renamed in a later patch to remove the try). Signed-off-by:
-
- 06 Oct, 2021 1 commit
-
-
Peter Zijlstra authored
When !SCHEDSTATS schedstat_enabled() is an unconditional 0 and the whole block doesn't exist, however GCC figures the scoped variable 'stats' is unused and complains about it. Upgrade the warning from -Wunused-variable to -Wunused-but-set-variable by writing it in two statements. This fixes the build because the new warning is in W=1. Given that whole if(0) {} thing, I don't feel motivated to change things overly much and quite strongly feel this is the compiler being daft. Fixes: cb3e971c435d ("sched: Make struct sched_statistics independent of fair sched class") Reported-by:Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by:
-
- 05 Oct, 2021 35 commits
-
-
Vincent Guittot authored
Since commit 89aafd67 ("sched/fair: Use prev instead of new target as recent_used_cpu"), p->recent_used_cpu is unconditionnaly set with prev. Fixes: 89aafd67 ("sched/fair: Use prev instead of new target as recent_used_cpu") Signed-off-by:
Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by:
Mel Gorman <mgorman@suse.de> Link: https://lkml.kernel.org/r/20210928103544.27489-1-vincent.guittot@linaro.org
-
Thomas Gleixner authored
Neither wq_worker_sleeping() nor io_wq_worker_sleeping() require to be invoked with preemption disabled: - The worker flag checks operations only need to be serialized against the worker thread itself. - The accounting and worker pool operations are serialized with locks. which means that disabling preemption has neither a reason nor a value. Remove it and update the stale comment. Signed-off-by:Thomas Gleixner <tglx@linutronix.de> Signed-off-by:
Lai Jiangshan <jiangshanlai@gmail.com> Reviewed-by:
Jens Axboe <axboe@kernel.dk> Link: https://lkml.kernel.org/r/8735pnafj7.ffs@tglx
-
Thomas Gleixner authored
Doing cleanups in the tail of schedule() is a latency punishment for the incoming task. The point of invoking kprobes_task_flush() for a dead task is that the instances are returned and cannot leak when __schedule() is kprobed. Move it into the delayed cleanup. Signed-off-by:
-
Thomas Gleixner authored
The queued remote wakeup mechanism has turned out to be suboptimal for RT enabled kernels. The maximum latencies go up by a factor of > 5x in certain scenarious. This is caused by either long wake lists or by a large number of TTWU IPIs which are processed back to back. Disable it for RT. Signed-off-by:
-
Thomas Gleixner authored
Batched task migrations are a source for large latencies as they keep the scheduler from running while processing the migrations. Limit the batch size to 8 instead of 32 when running on a RT enabled kernel. Signed-off-by:
-
Thomas Gleixner authored
mmdrop() is invoked from finish_task_switch() by the incoming task to drop the mm which was handed over by the previous task. mmdrop() can be quite expensive which prevents an incoming real-time task from getting useful work done. Provide mmdrop_sched() which maps to mmdrop() on !RT kernels. On RT kernels it delagates the eventually required invocation of __mmdrop() to RCU. Signed-off-by:
-
Shaokun Zhang authored
Make cookie functions static as these are no longer invoked directly by other code. No functional change intended. Signed-off-by:
Shaokun Zhang <zhangshaokun@hisilicon.com> Signed-off-by:
-
Ricardo Neri authored
When deciding to pull tasks in ASYM_PACKING, it is necessary not only to check for the idle state of the destination CPU, dst_cpu, but also of its SMT siblings. If dst_cpu is idle but its SMT siblings are busy, performance suffers if it pulls tasks from a medium priority CPU that does not have SMT siblings. Implement asym_smt_can_pull_tasks() to inspect the state of the SMT siblings of both dst_cpu and the CPUs in the candidate busiest group. Signed-off-by:
Ricardo Neri <ricardo.neri-calderon@linux.intel.com> Signed-off-by:
Joel Fernandes (Google) <joel@joelfernandes.org> Reviewed-by:
Len Brown <len.brown@intel.com> Reviewed-by:
-
Ricardo Neri authored
Create a separate function, sched_asym(). A subsequent changeset will introduce logic to deal with SMT in conjunction with asmymmetric packing. Such logic will need the statistics of the scheduling group provided as argument. Update them before calling sched_asym(). Co-developed-by:
-
Ricardo Neri authored
Before deciding to pull tasks when using asymmetric packing of tasks, on some architectures (e.g., x86) it is necessary to know not only the state of dst_cpu but also of its SMT siblings. The decision to classify a candidate busiest group as group_asym_packing is done in update_sg_lb_stats(). Give this function access to the scheduling domain statistics, which contains the statistics of the local group. Originally-by:
-
Ricardo Neri authored
sched_asmy_prefer() always returns false when called on the local group. By checking local_group, we can avoid additional checks and invoking sched_asmy_prefer() when it is not needed. No functional changes are introduced. Signed-off-by:
-
Ricardo Neri authored
There exist situations in which the load balance needs to know the properties of the CPUs in a scheduling group. When using asymmetric packing, for instance, the load balancer needs to know not only the state of dst_cpu but also of its SMT siblings, if any. Use the flags of the child scheduling domains to initialize scheduling group flags. This will reflect the properties of the CPUs in the group. A subsequent changeset will make use of these new flags. No functional changes are introduced. Originally-by:
-
Ricardo Neri authored
When scheduling, it is better to prefer a separate physical core rather than the SMT sibling of a high priority core. The existing formula to compute priorities takes such fact in consideration. There may exist, however, combinations of priorities (i.e., maximum frequencies) in which the priority of high-numbered SMT siblings of high-priority cores collides with the priority of low-numbered SMT siblings of low-priority cores. Consider for instance an SMT2 system with CPUs [0, 1] with priority 60 and [2, 3] with priority 30(CPUs in brackets are SMT siblings. In such a case, the resulting priorities would be [120, 60], [60, 30]. Thus, to ensure that CPU2 has higher priority than CPU1, divide the raw priority by the squared SMT iterator. The resulting priorities are [120, 30]. [60, 15]. Originally-by:
-
Sebastian Andrzej Siewior authored
With enabled threaded interrupts the nouveau driver reported the following: | Chain exists of: | &mm->mmap_lock#2 --> &device->mutex --> &cpuset_rwsem | | Possible unsafe locking scenario: | | CPU0 CPU1 | ---- ---- | lock(&cpuset_rwsem); | lock(&device->mutex); | lock(&cpuset_rwsem); | lock(&mm->mmap_lock#2); The device->mutex is nvkm_device::mutex. Unblocking the lockchain at `cpuset_rwsem' is probably the easiest thing to do. Move the priority reset to the start of the newly created thread. Fixes: 710da3c8 ("sched/core: Prevent race condition between cpuset and __sched_setscheduler()") Reported-by:
Mike Galbraith <efault@gmx.de> Signed-off-by:
Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by:
-
Frederic Weisbecker authored
Currently the boot defined preempt behaviour (aka dynamic preempt) selects full preemption by default when the "preempt=" boot parameter is omitted. However distros may rather want to default to either no preemption or voluntary preemption. To provide with this flexibility, make dynamic preemption a visible Kconfig option and adapt the preemption behaviour selected by the user to either static or dynamic preemption. Signed-off-by:
Frederic Weisbecker <frederic@kernel.org> Signed-off-by:
-
YueHaibing authored
These is no caller in tree since commit 523e979d ("sched/core: Use PELT for scale_rt_capacity()") Signed-off-by:
YueHaibing <yuehaibing@huawei.com> Signed-off-by:
-
Yafang Shao authored
After we make the struct sched_statistics and the helpers of it independent of fair sched class, we can easily use the schedstats facility for deadline sched class. The schedstat usage in DL sched class is similar with fair sched class, for example, fair deadline enqueue update_stats_enqueue_fair update_stats_enqueue_dl dequeue update_stats_dequeue_fair update_stats_dequeue_dl put_prev_task update_stats_wait_start update_stats_wait_start_dl set_next_task update_stats_wait_end update_stats_wait_end_dl The user can get the schedstats information in the same way in fair sched class. For example, fair deadline /proc/[pid]/sched /proc/[pid]/sched The output of a deadline task's schedstats as follows, $ cat /proc/69662/sched ... se.sum_exec_runtime : 3067.696449 se.nr_migrations : 0 sum_sleep_runtime : 720144.029661 sum_block_runtime : 0.547853 wait_start : 0.000000 sleep_start : 14131540.828955 block_start : 0.000000 sleep_max : 2999.974045 block_max : 0.283637 exec_max : 1.000269 slice_max : 0.000000 wait_max : 0.002217 wait_sum : 0.762179 wait_count : 733 iowait_sum : 0.547853 iowait_count : 3 nr_migrations_cold : 0 nr_failed_migrations_affine : 0 nr_failed_migrations_running : 0 nr_failed_migrations_hot : 0 nr_forced_migrations : 0 nr_wakeups : 246 nr_wakeups_sync : 2 nr_wakeups_migrate : 0 nr_wakeups_local : 244 nr_wakeups_remote : 2 nr_wakeups_affine : 0 nr_wakeups_affine_attempts : 0 nr_wakeups_passive : 0 nr_wakeups_idle : 0 ... The sched:sched_stat_{wait, sleep, iowait, blocked} tracepoints can be used to trace deadlline tasks as well. Signed-off-by:Yafang Shao <laoar.shao@gmail.com> Signed-off-by:
-
Yafang Shao authored
The runtime of a DL task has already been there, so we only need to add a tracepoint. One difference between fair task and DL task is that there is no vruntime in dl task. To reuse the sched_stat_runtime tracepoint, '0' is passed as vruntime for DL task. The output of this tracepoint for DL task as follows, top-36462 [047] d.h. 6083.452103: sched_stat_runtime: comm=top pid=36462 runtime=409898 [ns] vruntime=0 [ns] Signed-off-by: -
Yafang Shao authored
We want to measure the latency of RT tasks in our production environment with schedstats facility, but currently schedstats is only supported for fair sched class. This patch enable it for RT sched class as well. After we make the struct sched_statistics and the helpers of it independent of fair sched class, we can easily use the schedstats facility for RT sched class. The schedstat usage in RT sched class is similar with fair sched class, for example, fair RT enqueue update_stats_enqueue_fair update_stats_enqueue_rt dequeue update_stats_dequeue_fair update_stats_dequeue_rt put_prev_task update_stats_wait_start update_stats_wait_start_rt set_next_task update_stats_wait_end update_stats_wait_end_rt The user can get the schedstats information in the same way in fair sched class. For example, fair RT /proc/[pid]/sched /proc/[pid]/sched schedstats is not supported for RT group. The output of a RT task's schedstats as follows, $ cat /proc/10349/sched ... sum_sleep_runtime : 972.434535 sum_block_runtime : 960.433522 wait_start : 188510.871584 sleep_start : 0.000000 block_start : 0.000000 sleep_max : 12.001013 block_max : 952.660622 exec_max : 0.049629 slice_max : 0.000000 wait_max : 0.018538 wait_sum : 0.424340 wait_count : 49 iowait_sum : 956.495640 iowait_count : 24 nr_migrations_cold : 0 nr_failed_migrations_affine : 0 nr_failed_migrations_running : 0 nr_failed_migrations_hot : 0 nr_forced_migrations : 0 nr_wakeups : 49 nr_wakeups_sync : 0 nr_wakeups_migrate : 0 nr_wakeups_local : 49 nr_wakeups_remote : 0 nr_wakeups_affine : 0 nr_wakeups_affine_attempts : 0 nr_wakeups_passive : 0 nr_wakeups_idle : 0 ... The sched:sched_stat_{wait, sleep, iowait, blocked} tracepoints can be used to trace RT tasks as well. The output of these tracepoints for a RT tasks as follows, - runtime stress-10352 [004] d.h. 1035.382286: sched_stat_runtime: comm=stress pid=10352 runtime=995769 [ns] vruntime=0 [ns] [vruntime=0 means it is a RT task] - wait <idle>-0 [004] dN.. 1227.688544: sched_stat_wait: comm=stress pid=10352 delay=46849882 [ns] - blocked kworker/4:1-465 [004] dN.. 1585.676371: sched_stat_blocked: comm=stress pid=17194 delay=189963 [ns] - iowait kworker/4:1-465 [004] dN.. 1585.675330: sched_stat_iowait: comm=stress pid=17189 delay=182848 [ns] - sleep sleep-18194 [023] dN.. 1780.891840: sched_stat_sleep: comm=sleep.sh pid=17767 delay=1001160770 [ns] sleep-18196 [023] dN.. 1781.893208: sched_stat_sleep: comm=sleep.sh pid=17767 delay=1001161970 [ns] sleep-18197 [023] dN.. 1782.894544: sched_stat_sleep: comm=sleep.sh pid=17767 delay=1001128840 [ns] [ In sleep.sh, it sleeps 1 sec each time. ] [lkp@intel.com: reported build failure in earlier version] Signed-off-by: -
Yafang Shao authored
The runtime of a RT task has already been there, so we only need to add a tracepoint. One difference between fair task and RT task is that there is no vruntime in RT task. To reuse the sched_stat_runtime tracepoint, '0' is passed as vruntime for RT task. The output of this tracepoint for RT task as follows, stress-9748 [039] d.h. 113.519352: sched_stat_runtime: comm=stress pid=9748 runtime=997573 [ns] vruntime=0 [ns] stress-9748 [039] d.h. 113.520352: sched_stat_runtime: comm=stress pid=9748 runtime=997627 [ns] vruntime=0 [ns] stress-9748 [039] d.h. 113.521352: sched_stat_runtime: comm=stress pid=9748 runtime=998203 [ns] vruntime=0 [ns] Signed-off-by: -
Yafang Shao authored
Currently in schedstats we have sum_sleep_runtime and iowait_sum, but there's no metric to show how long the task is in D state. Once a task in D state, it means the task is blocked in the kernel, for example the task may be waiting for a mutex. The D state is more frequent than iowait, and it is more critital than S state. So it is worth to add a metric to measure it. Signed-off-by:
-
Yafang Shao authored
The original prototype of the schedstats helpers are update_stats_wait_*(struct cfs_rq *cfs_rq, struct sched_entity *se) The cfs_rq in these helpers is used to get the rq_clock, and the se is used to get the struct sched_statistics and the struct task_struct. In order to make these helpers available by all sched classes, we can pass the rq, sched_statistics and task_struct directly. Then the new helpers are update_stats_wait_*(struct rq *rq, struct task_struct *p, struct sched_statistics *stats) which are independent of fair sched class. To avoid vmlinux growing too large or introducing ovehead when !schedstat_enabled(), some new helpers after schedstat_enabled() are also introduced, Suggested by Mel. These helpers are in sched/stats.c, __update_stats_wait_*(struct rq *rq, struct task_struct *p, struct sched_statistics *stats) The size of vmlinux as follows, Before After Size of vmlinux 826308552 826304640 The size is a litte smaller as some functions are not inlined again after the change. I also compared the sched performance with 'perf bench sched pipe', suggested by Mel. The result as followsi (in usecs/op), Before After kernel.sched_schedstats=0 5.2~5.4 5.2~5.4 kernel.sched_schedstats=1 5.3~5.5 5.3~5.5 [These data is a little difference with the prev version, that is because my old test machine is destroyed so I have to use a new different test machine.] Almost no difference. No functional change. [lkp@intel.com: reported build failure in prev version] Signed-off-by: -
Yafang Shao authored
If we want to use the schedstats facility to trace other sched classes, we should make it independent of fair sched class. The struct sched_statistics is the schedular statistics of a task_struct or a task_group. So we can move it into struct task_struct and struct task_group to achieve the goal. After the patch, schestats are orgnized as follows, struct task_struct { ... struct sched_entity se; struct sched_rt_entity rt; struct sched_dl_entity dl; ... struct sched_statistics stats; ... }; Regarding the task group, schedstats is only supported for fair group sched, and a new struct sched_entity_stats is introduced, suggested by Peter - struct sched_entity_stats { struct sched_entity se; struct sched_statistics stats; } __no_randomize_layout; Then with the se in a task_group, we can easily get the stats. The sched_statistics members may be frequently modified when schedstats is enabled, in order to avoid impacting on random data which may in the same cacheline with them, the struct sched_statistics is defined as cacheline aligned. As this patch changes the core struct of scheduler, so I verified the performance it may impact on the scheduler with 'perf bench sched pipe', suggested by Mel. Below is the result, in which all the values are in usecs/op. Before After kernel.sched_schedstats=0 5.2~5.4 5.2~5.4 kernel.sched_schedstats=1 5.3~5.5 5.3~5.5 [These data is a little difference with the earlier version, that is because my old test machine is destroyed so I have to use a new different test machine.] Almost no impact on the sched performance. No functional change. [lkp@intel.com: reported build failure in earlier version] Signed-off-by: -
Yafang Shao authored
schedstat_enabled() has been already checked, so we can use __schedstat_set() directly. Signed-off-by:
-
Li Zhijian authored
Previously, 'make -C sched run_tests' will block forever when it occurs something wrong where the *selftests framework* is waiting for its child processes to exit. [root@iaas-rpma sched]# ./cs_prctl_test ## Create a thread/process/process group hiearchy Not a core sched system tid=74985, / tgid=74985 / pgid=74985: ffffffffffffffff Not a core sched system tid=74986, / tgid=74986 / pgid=74985: ffffffffffffffff Not a core sched system tid=74988, / tgid=74986 / pgid=74985: ffffffffffffffff Not a core sched system tid=74989, / tgid=74986 / pgid=74985: ffffffffffffffff Not a core sched system tid=74990, / tgid=74986 / pgid=74985: ffffffffffffffff Not a core sched system tid=74987, / tgid=74987 / pgid=74985: ffffffffffffffff Not a core sched system tid=74991, / tgid=74987 / pgid=74985: ffffffffffffffff Not a core sched system tid=74992, / tgid=74987 / pgid=74985: ffffffffffffffff Not a core sched system tid=74993, / tgid=74987 / pgid=74985: ffffffffffffffff Not a core sched system (268) FAILED: get_cs_cookie(0) == 0 ## Set a cookie on entire process group -1 = prctl(62, 1, 0, 2, 0) core_sched create failed -- PGID: Invalid argument (cs_prctl_test.c:272) - [root@iaas-rpma sched]# ps PID TTY TIME CMD 4605 pts/2 00:00:00 bash 74986 pts/2 00:00:00 cs_prctl_test 74987 pts/2 00:00:00 cs_prctl_test 74999 pts/2 00:00:00 ps Reported-by:kernel test robot <lkp@intel.com> Signed-off-by:
Li Zhijian <lizhijian@cn.fujitsu.com> Signed-off-by:
Chris Hyser <chris.hyser@oracle.com> Link: https://lore.kernel.org/r/20210902024333.75983-1-lizhijian@cn.fujitsu.com
-
Huaixin Chang authored
Basic description of usage and effect for CFS Bandwidth Control Burst. Co-developed-by:
Shanpei Chen <shanpeic@linux.alibaba.com> Signed-off-by:
Tianchen Ding <dtcccc@linux.alibaba.com> Signed-off-by:
Huaixin Chang <changhuaixin@linux.alibaba.com> Signed-off-by:
Daniel Jordan <daniel.m.jordan@oracle.com> Acked-by:
Tejun Heo <tj@kernel.org> Link: https://lore.kernel.org/r/20210830032215.16302-3-changhuaixin@linux.alibaba.com
-
Huaixin Chang authored
Two new statistics are introduced to show the internal of burst feature and explain why burst helps or not. nr_bursts: number of periods bandwidth burst occurs burst_time: cumulative wall-time (in nanoseconds) that any cpus has used above quota in respective periods Co-developed-by:
-
Josh Don authored
Give reduced sleeper credit to SCHED_IDLE entities. As a result, woken SCHED_IDLE entities will take longer to preempt normal entities. The benefit of this change is to make it less likely that a newly woken SCHED_IDLE entity will preempt a short-running normal entity before it blocks. We still give a small sleeper credit to SCHED_IDLE entities, so that idle<->idle competition retains some fairness. Example: With HZ=1000, spawned four threads affined to one cpu, one of which was set to SCHED_IDLE. Without this patch, wakeup latency for the SCHED_IDLE thread was ~1-2ms, with the patch the wakeup latency was ~5ms. Signed-off-by:
Josh Don <joshdon@google.com> Signed-off-by:
Jiang Biao <benbjiang@tencent.com> Link: https://lore.kernel.org/r/20210820010403.946838-5-joshdon@google.com
-
Josh Don authored
Use a small, non-scaled min granularity for SCHED_IDLE entities, when competing with normal entities. This reduces the latency of getting a normal entity back on cpu, at the expense of increased context switch frequency of SCHED_IDLE entities. The benefit of this change is to reduce the round-robin latency for normal entities when competing with a SCHED_IDLE entity. Example: on a machine with HZ=1000, spawned two threads, one of which is SCHED_IDLE, and affined to one cpu. Without this patch, the SCHED_IDLE thread runs for 4ms then waits for 1.4s. With this patch, it runs for 1ms and waits 340ms (as it round-robins with the other thread). Signed-off-by:
-
Josh Don authored
Adds cfs_rq->idle_nr_running, which accounts the number of idle entities directly enqueued on the cfs_rq. Signed-off-by:
-
Josh Don authored
/proc/uptime reports idle time by reading the CPUTIME_IDLE field from the per-cpu kcpustats. However, on NO_HZ systems, idle time is not continually updated on idle cpus, leading this value to appear incorrectly small. /proc/stat performs an accounting update when reading idle time; we can use the same approach for uptime. With this patch, /proc/stat and /proc/uptime now agree on idle time. Additionally, the following shows idle time tick up consistently on an idle machine: (while true; do cat /proc/uptime; sleep 1; done) | awk '{print $2-prev; prev=$2}' Reported-by:Luigi Rizzo <lrizzo@google.com> Signed-off-by:
Eric Dumazet <edumazet@google.com> Link: https://lkml.kernel.org/r/20210827165438.3280779-1-joshdon@google.com
-
Peter Zijlstra authored
Tao suggested a two-pass task selection to avoid the retry loop. Not only does it avoid the retry loop, it results in *much* simpler code. This also fixes an issue spotted by Josh Don where, for SMT3+, we can forget to update max on the first pass and get to do an extra round. Suggested-by:
Tao Zhou <tao.zhou@linux.dev> Signed-off-by:
Vineeth Pillai (Microsoft) <vineethrp@gmail.com> Link: https://lkml.kernel.org/r/YSS9+k1teA9oPEKl@hirez.programming.kicks-ass.net
-
Sebastian Andrzej Siewior authored
With PREEMPT_RT enabled all hrtimers callbacks will be invoked in softirq mode unless they are explicitly marked as HRTIMER_MODE_HARD. During boot kthread_bind() is used for the creation of per-CPU threads and then hangs in wait_task_inactive() if the ksoftirqd is not yet up and running. The hang disappeared since commit 26c7295b ("kthread: Do not preempt current task if it is going to call schedule()") but enabling function trace on boot reliably leads to the freeze on boot behaviour again. The timer in wait_task_inactive() can not be directly used by a user interface to abuse it and create a mass wake up of several tasks at the same time leading to long sections with disabled interrupts. Therefore it is safe to make the timer HRTIMER_MODE_REL_HARD. Switch the timer to HRTIMER_MODE_REL_HARD. Signed-off-by:
-
Valentin Schneider authored
Consider a system with some NOHZ-idle CPUs, such that nohz.idle_cpus_mask = S nohz.next_balance = T When a new CPU k goes NOHZ idle (nohz_balance_enter_idle()), we end up with: nohz.idle_cpus_mask = S \U {k} nohz.next_balance = T Note that the nohz.next_balance hasn't changed - it won't be updated until a NOHZ balance is triggered. This is problematic if the newly NOHZ idle CPU has an earlier rq.next_balance than the other NOHZ idle CPUs, IOW if: cpu_rq(k).next_balance < nohz.next_balance In such scenarios, the existing nohz.next_balance will prevent any NOHZ balance from happening, which itself will prevent nohz.next_balance from being updated to this new cpu_rq(k).next_balance. Unnecessary load balance delays of over 12ms caused by this were observed on an arm64 RB5 board. Use the new nohz.needs_update flag to mark the presence of newly-idle CPUs that need their rq->next_balance to be collated into nohz.next_balance. Trigger a NOHZ_NEXT_KICK when the flag is set. Signed-off-by:Valentin Schneider <valentin.schneider@arm.com> Signed-off-by:
-
Valentin Schneider authored
A following patch will trigger NOHZ idle balances as a means to update nohz.next_balance. Vincent noted that blocked load updates can have non-negligible overhead, which should be avoided if the intent is to only update nohz.next_balance. Add a new NOHZ balance kick flag, NOHZ_NEXT_KICK. Gate NOHZ blocked load update by the presence of NOHZ_STATS_KICK - currently all NOHZ balance kicks will have the NOHZ_STATS_KICK flag set, so no change in behaviour is expected. Suggested-by:
-
