forked from ~ljy/RK356X_SDK_RELEASE
blame | history | patch | commit | commitdiff
hc
2023-12-11 072de836f53be56a70cecf70b43ae43b7ce17376 
 kernel/kernel/sched/cputime.c
....@@ -1,8 +1,10 @@
1
+// SPDX-License-Identifier: GPL-2.0-only
12 /*
23  * Simple CPU accounting cgroup controller
34  */
45 #include <linux/cpufreq_times.h>
56 #include "sched.h"
7
+#include <trace/hooks/sched.h>
68 
79 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
810 
....@@ -18,6 +20,7 @@
1820  * compromise in place of having locks on each irq in account_system_time.
1921  */
2022 DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
23
+EXPORT_PER_CPU_SYMBOL_GPL(cpu_irqtime);
2124 
2225 static int sched_clock_irqtime;
2326 
....@@ -70,6 +73,8 @@
7073 		irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
7174 	else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
7275 		irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
76
+
77
+	trace_android_rvh_account_irq(curr, cpu, delta);
7378 }
7479 EXPORT_SYMBOL_GPL(irqtime_account_irq);
7580 
....@@ -361,7 +366,7 @@
361366  * softirq as those do not count in task exec_runtime any more.
362367  */
363368 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
364
-					 struct rq *rq, int ticks)
369
+					 int ticks)
365370 {
366371 	u64 other, cputime = TICK_NSEC * ticks;
367372 
....@@ -387,51 +392,48 @@
387392 		account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
388393 	} else if (user_tick) {
389394 		account_user_time(p, cputime);
390
-	} else if (p == rq->idle) {
395
+	} else if (p == this_rq()->idle) {
391396 		account_idle_time(cputime);
392397 	} else if (p->flags & PF_VCPU) { /* System time or guest time */
393398 		account_guest_time(p, cputime);
394399 	} else {
395400 		account_system_index_time(p, cputime, CPUTIME_SYSTEM);
396401 	}
402
+	trace_android_vh_irqtime_account_process_tick(p, this_rq(), user_tick, ticks);
397403 }
398404 
399405 static void irqtime_account_idle_ticks(int ticks)
400406 {
401
-	struct rq *rq = this_rq();
402
-
403
-	irqtime_account_process_tick(current, 0, rq, ticks);
407
+	irqtime_account_process_tick(current, 0, ticks);
404408 }
405409 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
406410 static inline void irqtime_account_idle_ticks(int ticks) { }
407411 static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
408
-						struct rq *rq, int nr_ticks) { }
412
+						int nr_ticks) { }
409413 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
410414 
411415 /*
412416  * Use precise platform statistics if available:
413417  */
414
-#ifdef CONFIG_VIRT_CPU_ACCOUNTING
418
+#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
419
+
415420 # ifndef __ARCH_HAS_VTIME_TASK_SWITCH
416
-void vtime_common_task_switch(struct task_struct *prev)
421
+void vtime_task_switch(struct task_struct *prev)
417422 {
418423 	if (is_idle_task(prev))
419424 		vtime_account_idle(prev);
420425 	else
421
-		vtime_account_system(prev);
426
+		vtime_account_kernel(prev);
422427 
423428 	vtime_flush(prev);
424429 	arch_vtime_task_switch(prev);
425430 }
426431 # endif
427
-#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
428432 
429
-
430
-#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
431433 /*
432434  * Archs that account the whole time spent in the idle task
433435  * (outside irq) as idle time can rely on this and just implement
434
- * vtime_account_system() and vtime_account_idle(). Archs that
436
+ * vtime_account_kernel() and vtime_account_idle(). Archs that
435437  * have other meaning of the idle time (s390 only includes the
436438  * time spent by the CPU when it's in low power mode) must override
437439  * vtime_account().
....@@ -442,7 +444,7 @@
442444 	if (!in_interrupt() && is_idle_task(tsk))
443445 		vtime_account_idle(tsk);
444446 	else
445
-		vtime_account_system(tsk);
447
+		vtime_account_kernel(tsk);
446448 }
447449 EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
448450 #endif /* __ARCH_HAS_VTIME_ACCOUNT */
....@@ -470,6 +472,7 @@
470472 	*ut = cputime.utime;
471473 	*st = cputime.stime;
472474 }
475
+EXPORT_SYMBOL_GPL(thread_group_cputime_adjusted);
473476 
474477 #else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
475478 
....@@ -481,13 +484,13 @@
481484 void account_process_tick(struct task_struct *p, int user_tick)
482485 {
483486 	u64 cputime, steal;
484
-	struct rq *rq = this_rq();
485487 
486
-	if (vtime_accounting_cpu_enabled())
488
+	if (vtime_accounting_enabled_this_cpu())
487489 		return;
490
+	trace_android_vh_account_task_time(p, this_rq(), user_tick);
488491 
489492 	if (sched_clock_irqtime) {
490
-		irqtime_account_process_tick(p, user_tick, rq, 1);
493
+		irqtime_account_process_tick(p, user_tick, 1);
491494 		return;
492495 	}
493496 
....@@ -501,7 +504,7 @@
501504 
502505 	if (user_tick)
503506 		account_user_time(p, cputime);
504
-	else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
507
+	else if ((p != this_rq()->idle) || (irq_count() != HARDIRQ_OFFSET))
505508 		account_system_time(p, HARDIRQ_OFFSET, cputime);
506509 	else
507510 		account_idle_time(cputime);
....@@ -528,50 +531,6 @@
528531 
529532 	cputime -= steal;
530533 	account_idle_time(cputime);
531
-}
532
-
533
-/*
534
- * Perform (stime * rtime) / total, but avoid multiplication overflow by
535
- * loosing precision when the numbers are big.
536
- */
537
-static u64 scale_stime(u64 stime, u64 rtime, u64 total)
538
-{
539
-	u64 scaled;
540
-
541
-	for (;;) {
542
-		/* Make sure "rtime" is the bigger of stime/rtime */
543
-		if (stime > rtime)
544
-			swap(rtime, stime);
545
-
546
-		/* Make sure 'total' fits in 32 bits */
547
-		if (total >> 32)
548
-			goto drop_precision;
549
-
550
-		/* Does rtime (and thus stime) fit in 32 bits? */
551
-		if (!(rtime >> 32))
552
-			break;
553
-
554
-		/* Can we just balance rtime/stime rather than dropping bits? */
555
-		if (stime >> 31)
556
-			goto drop_precision;
557
-
558
-		/* We can grow stime and shrink rtime and try to make them both fit */
559
-		stime <<= 1;
560
-		rtime >>= 1;
561
-		continue;
562
-
563
-drop_precision:
564
-		/* We drop from rtime, it has more bits than stime */
565
-		rtime >>= 1;
566
-		total >>= 1;
567
-	}
568
-
569
-	/*
570
-	 * Make sure gcc understands that this is a 32x32->64 multiply,
571
-	 * followed by a 64/32->64 divide.
572
-	 */
573
-	scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
574
-	return scaled;
575534 }
576535 
577536 /*
....@@ -633,7 +592,7 @@
633592 		goto update;
634593 	}
635594 
636
-	stime = scale_stime(stime, rtime, stime + utime);
595
+	stime = mul_u64_u64_div_u64(stime, rtime, stime + utime);
637596 
638597 update:
639598 	/*
....@@ -684,6 +643,8 @@
684643 	thread_group_cputime(p, &cputime);
685644 	cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
686645 }
646
+EXPORT_SYMBOL_GPL(thread_group_cputime_adjusted);
647
+
687648 #endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
688649 
689650 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
....@@ -717,8 +678,8 @@
717678 	return delta - other;
718679 }
719680 
720
-static void __vtime_account_system(struct task_struct *tsk,
721
-				   struct vtime *vtime)
681
+static void vtime_account_system(struct task_struct *tsk,
682
+				 struct vtime *vtime)
722683 {
723684 	vtime->stime += get_vtime_delta(vtime);
724685 	if (vtime->stime >= TICK_NSEC) {
....@@ -737,7 +698,17 @@
737698 	}
738699 }
739700 
740
-void vtime_account_system(struct task_struct *tsk)
701
+static void __vtime_account_kernel(struct task_struct *tsk,
702
+				   struct vtime *vtime)
703
+{
704
+	/* We might have scheduled out from guest path */
705
+	if (vtime->state == VTIME_GUEST)
706
+		vtime_account_guest(tsk, vtime);
707
+	else
708
+		vtime_account_system(tsk, vtime);
709
+}
710
+
711
+void vtime_account_kernel(struct task_struct *tsk)
741712 {
742713 	struct vtime *vtime = &tsk->vtime;
743714 
....@@ -745,11 +716,7 @@
745716 		return;
746717 
747718 	write_seqcount_begin(&vtime->seqcount);
748
-	/* We might have scheduled out from guest path */
749
-	if (tsk->flags & PF_VCPU)
750
-		vtime_account_guest(tsk, vtime);
751
-	else
752
-		__vtime_account_system(tsk, vtime);
719
+	__vtime_account_kernel(tsk, vtime);
753720 	write_seqcount_end(&vtime->seqcount);
754721 }
755722 
....@@ -758,7 +725,7 @@
758725 	struct vtime *vtime = &tsk->vtime;
759726 
760727 	write_seqcount_begin(&vtime->seqcount);
761
-	__vtime_account_system(tsk, vtime);
728
+	vtime_account_system(tsk, vtime);
762729 	vtime->state = VTIME_USER;
763730 	write_seqcount_end(&vtime->seqcount);
764731 }
....@@ -788,8 +755,9 @@
788755 	 * that can thus safely catch up with a tickless delta.
789756 	 */
790757 	write_seqcount_begin(&vtime->seqcount);
791
-	__vtime_account_system(tsk, vtime);
758
+	vtime_account_system(tsk, vtime);
792759 	tsk->flags |= PF_VCPU;
760
+	vtime->state = VTIME_GUEST;
793761 	write_seqcount_end(&vtime->seqcount);
794762 }
795763 EXPORT_SYMBOL_GPL(vtime_guest_enter);
....@@ -801,6 +769,7 @@
801769 	write_seqcount_begin(&vtime->seqcount);
802770 	vtime_account_guest(tsk, vtime);
803771 	tsk->flags &= ~PF_VCPU;
772
+	vtime->state = VTIME_SYS;
804773 	write_seqcount_end(&vtime->seqcount);
805774 }
806775 EXPORT_SYMBOL_GPL(vtime_guest_exit);
....@@ -810,19 +779,30 @@
810779 	account_idle_time(get_vtime_delta(&tsk->vtime));
811780 }
812781 
813
-void arch_vtime_task_switch(struct task_struct *prev)
782
+void vtime_task_switch_generic(struct task_struct *prev)
814783 {
815784 	struct vtime *vtime = &prev->vtime;
816785 
817786 	write_seqcount_begin(&vtime->seqcount);
787
+	if (vtime->state == VTIME_IDLE)
788
+		vtime_account_idle(prev);
789
+	else
790
+		__vtime_account_kernel(prev, vtime);
818791 	vtime->state = VTIME_INACTIVE;
792
+	vtime->cpu = -1;
819793 	write_seqcount_end(&vtime->seqcount);
820794 
821795 	vtime = &current->vtime;
822796 
823797 	write_seqcount_begin(&vtime->seqcount);
824
-	vtime->state = VTIME_SYS;
798
+	if (is_idle_task(current))
799
+		vtime->state = VTIME_IDLE;
800
+	else if (current->flags & PF_VCPU)
801
+		vtime->state = VTIME_GUEST;
802
+	else
803
+		vtime->state = VTIME_SYS;
825804 	vtime->starttime = sched_clock();
805
+	vtime->cpu = smp_processor_id();
826806 	write_seqcount_end(&vtime->seqcount);
827807 }
828808 
....@@ -833,8 +813,9 @@
833813 
834814 	local_irq_save(flags);
835815 	write_seqcount_begin(&vtime->seqcount);
836
-	vtime->state = VTIME_SYS;
816
+	vtime->state = VTIME_IDLE;
837817 	vtime->starttime = sched_clock();
818
+	vtime->cpu = cpu;
838819 	write_seqcount_end(&vtime->seqcount);
839820 	local_irq_restore(flags);
840821 }
....@@ -852,7 +833,7 @@
852833 		seq = read_seqcount_begin(&vtime->seqcount);
853834 
854835 		gtime = t->gtime;
855
-		if (vtime->state == VTIME_SYS && t->flags & PF_VCPU)
836
+		if (vtime->state == VTIME_GUEST)
856837 			gtime += vtime->gtime + vtime_delta(vtime);
857838 
858839 	} while (read_seqcount_retry(&vtime->seqcount, seq));
....@@ -883,20 +864,233 @@
883864 		*utime = t->utime;
884865 		*stime = t->stime;
885866 
886
-		/* Task is sleeping, nothing to add */
887
-		if (vtime->state == VTIME_INACTIVE || is_idle_task(t))
867
+		/* Task is sleeping or idle, nothing to add */
868
+		if (vtime->state < VTIME_SYS)
888869 			continue;
889870 
890871 		delta = vtime_delta(vtime);
891872 
892873 		/*
893
-		 * Task runs either in user or kernel space, add pending nohz time to
894
-		 * the right place.
874
+		 * Task runs either in user (including guest) or kernel space,
875
+		 * add pending nohz time to the right place.
895876 		 */
896
-		if (vtime->state == VTIME_USER || t->flags & PF_VCPU)
897
-			*utime += vtime->utime + delta;
898
-		else if (vtime->state == VTIME_SYS)
877
+		if (vtime->state == VTIME_SYS)
899878 			*stime += vtime->stime + delta;
879
+		else
880
+			*utime += vtime->utime + delta;
900881 	} while (read_seqcount_retry(&vtime->seqcount, seq));
901882 }
883
+
884
+static int vtime_state_fetch(struct vtime *vtime, int cpu)
885
+{
886
+	int state = READ_ONCE(vtime->state);
887
+
888
+	/*
889
+	 * We raced against a context switch, fetch the
890
+	 * kcpustat task again.
891
+	 */
892
+	if (vtime->cpu != cpu && vtime->cpu != -1)
893
+		return -EAGAIN;
894
+
895
+	/*
896
+	 * Two possible things here:
897
+	 * 1) We are seeing the scheduling out task (prev) or any past one.
898
+	 * 2) We are seeing the scheduling in task (next) but it hasn't
899
+	 *    passed though vtime_task_switch() yet so the pending
900
+	 *    cputime of the prev task may not be flushed yet.
901
+	 *
902
+	 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
903
+	 */
904
+	if (state == VTIME_INACTIVE)
905
+		return -EAGAIN;
906
+
907
+	return state;
908
+}
909
+
910
+static u64 kcpustat_user_vtime(struct vtime *vtime)
911
+{
912
+	if (vtime->state == VTIME_USER)
913
+		return vtime->utime + vtime_delta(vtime);
914
+	else if (vtime->state == VTIME_GUEST)
915
+		return vtime->gtime + vtime_delta(vtime);
916
+	return 0;
917
+}
918
+
919
+static int kcpustat_field_vtime(u64 *cpustat,
920
+				struct task_struct *tsk,
921
+				enum cpu_usage_stat usage,
922
+				int cpu, u64 *val)
923
+{
924
+	struct vtime *vtime = &tsk->vtime;
925
+	unsigned int seq;
926
+
927
+	do {
928
+		int state;
929
+
930
+		seq = read_seqcount_begin(&vtime->seqcount);
931
+
932
+		state = vtime_state_fetch(vtime, cpu);
933
+		if (state < 0)
934
+			return state;
935
+
936
+		*val = cpustat[usage];
937
+
938
+		/*
939
+		 * Nice VS unnice cputime accounting may be inaccurate if
940
+		 * the nice value has changed since the last vtime update.
941
+		 * But proper fix would involve interrupting target on nice
942
+		 * updates which is a no go on nohz_full (although the scheduler
943
+		 * may still interrupt the target if rescheduling is needed...)
944
+		 */
945
+		switch (usage) {
946
+		case CPUTIME_SYSTEM:
947
+			if (state == VTIME_SYS)
948
+				*val += vtime->stime + vtime_delta(vtime);
949
+			break;
950
+		case CPUTIME_USER:
951
+			if (task_nice(tsk) <= 0)
952
+				*val += kcpustat_user_vtime(vtime);
953
+			break;
954
+		case CPUTIME_NICE:
955
+			if (task_nice(tsk) > 0)
956
+				*val += kcpustat_user_vtime(vtime);
957
+			break;
958
+		case CPUTIME_GUEST:
959
+			if (state == VTIME_GUEST && task_nice(tsk) <= 0)
960
+				*val += vtime->gtime + vtime_delta(vtime);
961
+			break;
962
+		case CPUTIME_GUEST_NICE:
963
+			if (state == VTIME_GUEST && task_nice(tsk) > 0)
964
+				*val += vtime->gtime + vtime_delta(vtime);
965
+			break;
966
+		default:
967
+			break;
968
+		}
969
+	} while (read_seqcount_retry(&vtime->seqcount, seq));
970
+
971
+	return 0;
972
+}
973
+
974
+u64 kcpustat_field(struct kernel_cpustat *kcpustat,
975
+		   enum cpu_usage_stat usage, int cpu)
976
+{
977
+	u64 *cpustat = kcpustat->cpustat;
978
+	u64 val = cpustat[usage];
979
+	struct rq *rq;
980
+	int err;
981
+
982
+	if (!vtime_accounting_enabled_cpu(cpu))
983
+		return val;
984
+
985
+	rq = cpu_rq(cpu);
986
+
987
+	for (;;) {
988
+		struct task_struct *curr;
989
+
990
+		rcu_read_lock();
991
+		curr = rcu_dereference(rq->curr);
992
+		if (WARN_ON_ONCE(!curr)) {
993
+			rcu_read_unlock();
994
+			return cpustat[usage];
995
+		}
996
+
997
+		err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
998
+		rcu_read_unlock();
999
+
1000
+		if (!err)
1001
+			return val;
1002
+
1003
+		cpu_relax();
1004
+	}
1005
+}
1006
+EXPORT_SYMBOL_GPL(kcpustat_field);
1007
+
1008
+static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
1009
+				    const struct kernel_cpustat *src,
1010
+				    struct task_struct *tsk, int cpu)
1011
+{
1012
+	struct vtime *vtime = &tsk->vtime;
1013
+	unsigned int seq;
1014
+
1015
+	do {
1016
+		u64 *cpustat;
1017
+		u64 delta;
1018
+		int state;
1019
+
1020
+		seq = read_seqcount_begin(&vtime->seqcount);
1021
+
1022
+		state = vtime_state_fetch(vtime, cpu);
1023
+		if (state < 0)
1024
+			return state;
1025
+
1026
+		*dst = *src;
1027
+		cpustat = dst->cpustat;
1028
+
1029
+		/* Task is sleeping, dead or idle, nothing to add */
1030
+		if (state < VTIME_SYS)
1031
+			continue;
1032
+
1033
+		delta = vtime_delta(vtime);
1034
+
1035
+		/*
1036
+		 * Task runs either in user (including guest) or kernel space,
1037
+		 * add pending nohz time to the right place.
1038
+		 */
1039
+		if (state == VTIME_SYS) {
1040
+			cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1041
+		} else if (state == VTIME_USER) {
1042
+			if (task_nice(tsk) > 0)
1043
+				cpustat[CPUTIME_NICE] += vtime->utime + delta;
1044
+			else
1045
+				cpustat[CPUTIME_USER] += vtime->utime + delta;
1046
+		} else {
1047
+			WARN_ON_ONCE(state != VTIME_GUEST);
1048
+			if (task_nice(tsk) > 0) {
1049
+				cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1050
+				cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1051
+			} else {
1052
+				cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1053
+				cpustat[CPUTIME_USER] += vtime->gtime + delta;
1054
+			}
1055
+		}
1056
+	} while (read_seqcount_retry(&vtime->seqcount, seq));
1057
+
1058
+	return 0;
1059
+}
1060
+
1061
+void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
1062
+{
1063
+	const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1064
+	struct rq *rq;
1065
+	int err;
1066
+
1067
+	if (!vtime_accounting_enabled_cpu(cpu)) {
1068
+		*dst = *src;
1069
+		return;
1070
+	}
1071
+
1072
+	rq = cpu_rq(cpu);
1073
+
1074
+	for (;;) {
1075
+		struct task_struct *curr;
1076
+
1077
+		rcu_read_lock();
1078
+		curr = rcu_dereference(rq->curr);
1079
+		if (WARN_ON_ONCE(!curr)) {
1080
+			rcu_read_unlock();
1081
+			*dst = *src;
1082
+			return;
1083
+		}
1084
+
1085
+		err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1086
+		rcu_read_unlock();
1087
+
1088
+		if (!err)
1089
+			return;
1090
+
1091
+		cpu_relax();
1092
+	}
1093
+}
1094
+EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1095
+
9021096 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */