@@ -2542,6 +2542,9 @@ try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
goto out;
}
+ if (is_per_cpu_kthread(p))
+ wake_flags |= WF_KTHREAD;
+
/*
* If we are going to wake up a thread waiting for CONDITION we
* need to ensure that CONDITION=1 done by the caller can not be
@@ -6716,7 +6716,7 @@ static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_
find_matching_se(&se, &pse);
update_curr(cfs_rq_of(se));
BUG_ON(!pse);
- if (wakeup_preempt_entity(se, pse) == 1) {
+ if (wakeup_preempt_entity(se, pse) >= !(wake_flags & WF_KTHREAD)) {
/*
* Bias pick_next to pick the sched entity that is
* triggering this preemption.
@@ -1643,7 +1643,8 @@ static inline int task_on_rq_migrating(struct task_struct *p)
*/
#define WF_SYNC 0x01 /* Waker goes to sleep after wakeup */
#define WF_FORK 0x02 /* Child wakeup after fork */
-#define WF_MIGRATED 0x4 /* Internal use, task got migrated */
+#define WF_MIGRATED 0x04 /* Internal use, task got migrated */
+#define WF_KTHREAD 0x08 /* Per CPU Kthread */
/*
* To aid in avoiding the subversion of "niceness" due to uneven distribution
A running task can wake-up a per CPU bound kthread on the same CPU. If the current running task doesn't yield the CPU before the next load balance operation, the scheduler would detect load imbalance and try to balance the load. However this load balance would fail as the waiting task is CPU bound, while the running task cannot be moved by the regular load balancer. Finally the active load balancer would kick in and move the task to a different CPU/Core. Moving the task to a different CPU/core can lead to loss in cache affinity leading to poor performance. This is more prone to happen if the current running task is CPU intensive and the sched_wake_up_granularity is set to larger value. When the sched_wake_up_granularity was relatively small, it was observed that the bound thread would complete before the load balancer would have chosen to move the cache hot task to a different CPU. To deal with this situation, the current running task would yield to a per CPU bound kthread, provided kthread is not CPU intensive. /pboffline/hwcct_prg_old/lib/fsperf -t overwrite --noclean -f 5g -b 4k /pboffline (With sched_wake_up_granularity set to 15ms) Performance counter stats for 'system wide' (5 runs): event v5.4 v5.4 + patch(v3) probe:active_load_balance_cpu_stop 1,919 ( +- 2.89% ) 4 ( +- 20.48% ) sched:sched_waking 441,535 ( +- 0.17% ) 914,630 ( +- 0.18% ) sched:sched_wakeup 441,533 ( +- 0.17% ) 914,630 ( +- 0.18% ) sched:sched_wakeup_new 2,436 ( +- 8.08% ) 545 ( +- 4.02% ) sched:sched_switch 797,007 ( +- 0.26% ) 1,490,261 ( +- 0.10% ) sched:sched_migrate_task 20,998 ( +- 1.04% ) 2,492 ( +- 11.56% ) sched:sched_process_free 2,436 ( +- 7.90% ) 526 ( +- 3.65% ) sched:sched_process_exit 2,451 ( +- 7.85% ) 546 ( +- 4.06% ) sched:sched_wait_task 7 ( +- 21.20% ) 1 ( +- 40.82% ) sched:sched_process_wait 3,951 ( +- 9.14% ) 854 ( +- 5.33% ) sched:sched_process_fork 2,435 ( +- 8.09% ) 545 ( +- 3.96% ) sched:sched_process_exec 1,023 ( +- 12.21% ) 205 ( +- 5.13% ) sched:sched_wake_idle_without_ipi 187,794 ( +- 1.14% ) 353,579 ( +- 0.42% ) Elasped time in seconds 289.43 +- 1.42 ( +- 0.49% ) 72.7318 +- 0.0545 ( +- 0.07% ) Throughput results v5.4 Trigger time:................... 0.842679 s (Throughput: 6075.86 MB/s) Asynchronous submit time:....... 1.0184 s (Throughput: 5027.49 MB/s) Synchronous submit time:........ 0 s (Throughput: 0 MB/s) I/O time:....................... 263.17 s (Throughput: 19.455 MB/s) Ratio trigger time to I/O time:.0.00320202 v5.4 + patch(v3) Trigger time:................... 0.852413 s (Throughput: 6006.47 MB/s) Asynchronous submit time:....... 0.773043 s (Throughput: 6623.17 MB/s) Synchronous submit time:........ 0 s (Throughput: 0 MB/s) I/O time:....................... 44.341 s (Throughput: 115.468 MB/s) Ratio trigger time to I/O time:. 0.019224 (With sched_wake_up_granularity set to 4ms) Performance counter stats for 'system wide' (5 runs): event v5.4 v5.4 + patch(v3) probe:active_load_balance_cpu_stop 6 ( +- 6.03% ) 5 ( +- 15.04% ) sched:sched_waking 899,880 ( +- 0.38% ) 912,625 ( +- 0.41% ) sched:sched_wakeup 899,878 ( +- 0.38% ) 912,624 ( +- 0.41% ) sched:sched_wakeup_new 622 ( +- 11.95% ) 550 ( +- 3.85% ) sched:sched_switch 1,458,214 ( +- 0.40% ) 1,489,032 ( +- 0.41% ) sched:sched_migrate_task 3,120 ( +- 10.00% ) 2,524 ( +- 5.54% ) sched:sched_process_free 608 ( +- 12.18% ) 528 ( +- 3.89% ) sched:sched_process_exit 623 ( +- 11.91% ) 550 ( +- 3.79% ) sched:sched_wait_task 1 ( +- 31.18% ) 1 ( +- 66.67% ) sched:sched_process_wait 998 ( +- 13.22% ) 867 ( +- 4.41% ) sched:sched_process_fork 622 ( +- 11.95% ) 550 ( +- 3.88% ) sched:sched_process_exec 242 ( +- 13.81% ) 208 ( +- 4.57% ) sched:sched_wake_idle_without_ipi 349,165 ( +- 0.35% ) 352,443 ( +- 0.21% ) Elasped time in seconds 72.8560 +- 0.0768 ( +- 0.11% ) 72.5523 +- 0.0725 ( +- 0.10% ) Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> --- Changelog: v1 : http://lore.kernel.org/lkml/20191209165122.GA27229@linux.vnet.ibm.com v2 : http://lore.kernel.org/lkml/20191210054330.GF27253@linux.vnet.ibm.com v1->v2: Pass the the right params to try_to_wake_up as correctly pointed out by Dave Chinner v2->v3: Suggestions from Peter Zijlstra including using vtime over context switch and detect per-cpu-kthread in try_to_wake_up kernel/sched/core.c | 3 +++ kernel/sched/fair.c | 2 +- kernel/sched/sched.h | 3 ++- 3 files changed, 6 insertions(+), 2 deletions(-)