commit 4117cebf1a9f ("psi: Optimize task switch inside shared cgroups")
defer prev task sleep handling to psi_task_switch(), so we don't need
to clear and set TSK_ONCPU state for common cgroups.
A
|
B
/ \
C D
/ \
prev next
After that commit psi_task_switch() do:
1. psi_group_change(next, .set=TSK_ONCPU) for D
2. psi_group_change(prev, .clear=TSK_ONCPU | TSK_RUNNING) for C
3. psi_group_change(prev, .clear=TSK_RUNNING) for B, A
But there is a limitation "prev->psi_flags == next->psi_flags" that
if not satisfied, will make this cgroups optimization unusable for both
sleep switch or running switch cases. For example:
prev->in_memstall != next->in_memstall when sleep switch:
1. psi_group_change(next, .set=TSK_ONCPU) for D, B, A
2. psi_group_change(prev, .clear=TSK_ONCPU | TSK_RUNNING) for C, B, A
prev->in_memstall != next->in_memstall when running switch:
1. psi_group_change(next, .set=TSK_ONCPU) for D, B, A
2. psi_group_change(prev, .clear=TSK_ONCPU) for C, B, A
The reason why this limitation exist is that we consider a group is
PSI_MEM_FULL if the CPU is actively reclaiming and nothing productive
could run even if it were runnable. So when CPU curr changed from prev
to next and their in_memstall status is different, we have to change
PSI_MEM_FULL status for their common cgroups.
This patch remove this limitation by making psi_group_change() change
PSI_MEM_FULL status depend on CPU curr->in_memstall status.
Signed-off-by: Chengming Zhou <zhouchengming@xxxxxxxxxxxxx>
---
kernel/sched/psi.c | 8 ++------
1 file changed, 2 insertions(+), 6 deletions(-)
diff --git a/kernel/sched/psi.c b/kernel/sched/psi.c
index 77d53c03a76f..26c03bd56b9c 100644
--- a/kernel/sched/psi.c
+++ b/kernel/sched/psi.c
@@ -820,8 +820,6 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
u64 now = cpu_clock(cpu);
if (next->pid) {
- bool identical_state;
-
psi_flags_change(next, 0, TSK_ONCPU);
/*
* When switching between tasks that have an identical
@@ -829,11 +827,9 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
* we reach the first common ancestor. Iterate @next's
* ancestors only until we encounter @prev's ONCPU.
*/
- identical_state = prev->psi_flags == next->psi_flags;
iter = NULL;
while ((group = iterate_groups(next, &iter))) {
- if (identical_state &&
- per_cpu_ptr(group->pcpu, cpu)->tasks[NR_ONCPU]) {
+ if (per_cpu_ptr(group->pcpu, cpu)->tasks[NR_ONCPU]) {
common = group;
break;
}
@@ -880,7 +876,7 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
* TSK_ONCPU is handled up to the common ancestor. If we're tasked
* with dequeuing too, finish that for the rest of the hierarchy.
*/
- if (sleep) {
+ if (sleep || unlikely(prev->in_memstall != next->in_memstall)) {
clear &= ~TSK_ONCPU;
for (; group; group = iterate_groups(prev, &iter))
psi_group_change(group, cpu, clear, set, now, wake_clock);
--
2.37.2
