sched: fix unfairness when upgrade weight
PATCH 1/1 sched: fix unfairness when upgrade weight
When two or more processes upgrade their priority, unfairness will happen, part of them may get all cpu-usage, and the others cannot be scheduled to run for a long time.
example:
# (create 2 processes and set affinity to cpu#0)
# renice 19 pid1 pid2 # renice -19 pid1 pid2
step3 upgrade the 2 processes' weight, these 2 processes should share the cpu#0 as soon as possible after step3, and each of them should get 50% cpu-usage. But sometimes one of them gets all cpu-usage for tens of seconds before the other is schedulered to run.
fair-group example:
# mkdir 1 2 (create 2 fair-groups) # (create 2 processes and set affinity to cpu#0) # echo pid1 > 1/tasks ; echo pid2 > 2/tasks # echo 2 > 1/cpu.shares ; echo 2 > 2/cpu.shares # echo $((2**18)) > 1/cpu.shares ; echo $((2**18)) > 2/cpu.shares
The reason why such unfairness happened:
When a sched_entity is running, if its weight is low, its vruntime increases by a large value every time and if its weight is high, its vruntime increases by a small value.
So when the two sched_entity's weight is low, they will still fairness even if difference of their vruntime is large, but if their weight are upgraded, this large difference of vruntime will bring unfairness. Because it will cost the process to spend a lot of time to catch up the huge difference.
example:
se1's vruntime se2's vruntime
1000M (R) 1020M
(assume vruntime is increases by about 50M every time)
(R) 1050M 1020M
1050M (R) 1070M
(R) 1100M 1070M
1100M (R) 1120M
(fairness, even if difference of their vruntime is large) (upgrade their weight, vruntime is increases by about 10K)
(R) 1100M+10K 1120M (R) 1100M+20K 1120M (R) 1100M+30K 1120M (R) 1100M+40K 1120M (R) 1100M+50K 1120M
(se1 gets all cpu-usage for long time (mybe about tens of seconds)) (unfairness, difference=20M is too large for new weight)
This patch fixes this bug by tuning the vruntime of weight-upgraded sched entities, just like waking up a task. the new vruntime will be
cfs_rq->min_vruntime + sched_vslice();
--- kernel/sched.c | 16 +++++++++------- kernel/sched_fair.c | 9 +++++++++ 2 files changed, 18 insertions(+), 7 deletions(-)
kernel/sched.c +9 -7
@@ -5096,12 +5096,8 @@ void set_user_nice(struct task_struct *p, long nice)if (on_rq) {enqueue_task (rq, p,0 ); -/* -* If the task increased its priority or is running and -* lowered its priority, then reschedule its CPU: -*/ -if (delta <0 || (delta >0 &&task_running (rq, p))) -resched_task (rq->curr); + p->sched_class->prio_changed (rq, p, old_prio, +task_running (rq, p)); } out_unlock:task_rq_unlock (rq, &flags);
@@ -8929,16 +8925,22 @@ static void __set_se_shares(struct sched_entity *se, unsigned long shares) {struct cfs_rq *cfs_rq = se->cfs_rq;int on_rq; +unsigned long old_weight; on_rq = se->on_rq;if (on_rq)dequeue_entity (cfs_rq, se,0 ); + old_weight = se->load.weight; se->load.weight = shares; se->load.inv_weight =0 ; -if (on_rq) +if (on_rq) { +if (se->load.weight > old_weight) + se->vruntime = cfs_rq->min_vruntime + +sched_vslice (cfs_rq, se);enqueue_entity (cfs_rq, se,0 ); + } }static void set_se_shares (struct sched_entity *se,unsigned long shares)
kernel/sched_fair.c +9 -0
@@ -1690,6 +1690,15 @@ static void task_new_fair(struct rq *rq, struct task_struct *p)static void prio_changed_fair (struct rq *rq,struct task_struct *p,int oldprio,int running) { +struct cfs_rq *cfs_rq =task_cfs_rq (p); +struct sched_entity *se = &p->se; +int on_rq = se->on_rq; + +if (p->prio < oldprio && on_rq) { +dequeue_entity (cfs_rq, se,0 ); + se->vruntime = cfs_rq->min_vruntime +sched_vslice (cfs_rq, se); +enqueue_entity (cfs_rq, se,0 ); + }/* * Reschedule if we are currently running on this runqueue and * our priority decreased, or if we are not currently running on
Signature · 1 line
1.6.0.3
Quoted 3 lines · Miao Xie
This patch fixes this bug by tuning the vruntime of weight-upgraded sched entities, just like waking up a task. the new vruntime will be
cfs_rq->min_vruntime + sched_vslice();
I really don't like that.
Better would be to scale with min_vruntime, which would at least approximate the lag somewhat.
Best is to compute the actual lag, but that might just not be worth the extra overhead.
http://programming.kicks-ass.net/kernel-patches/sched-avg_vruntime/
Quoted 55 lines
--- kernel/sched.c | 16 +++++++++------- kernel/sched_fair.c | 9 +++++++++ 2 files changed, 18 insertions(+), 7 deletions(-)kernel/sched.c +9
-7@@ -5096,12 +5096,8 @@ void set_user_nice(struct task_struct *p, long nice)if (on_rq) {enqueue_task (rq, p,0 ); -/* -* If the task increased its priority or is running and -* lowered its priority, then reschedule its CPU: -*/ -if (delta <0 || (delta >0 &&task_running (rq, p))) -resched_task (rq->curr); + p->sched_class->prio_changed (rq, p, old_prio, +task_running (rq, p)); } out_unlock:task_rq_unlock (rq, &flags);@@ -8929,16 +8925,22 @@ static void __set_se_shares(struct sched_entity *se, unsigned long shares) {struct cfs_rq *cfs_rq = se->cfs_rq;int on_rq; +unsigned long old_weight; on_rq = se->on_rq;if (on_rq)dequeue_entity (cfs_rq, se,0 ); + old_weight = se->load.weight; se->load.weight = shares; se->load.inv_weight =0 ; -if (on_rq) +if (on_rq) { +if (se->load.weight > old_weight) + se->vruntime = cfs_rq->min_vruntime + +sched_vslice (cfs_rq, se);enqueue_entity (cfs_rq, se,0 ); + } }static void set_se_shares (struct sched_entity *se,unsigned long shares)kernel/sched_fair.c +0
-0@@ -1690,6 +1690,15 @@ static void task_new_fair(struct rq *rq, struct task_struct *p)static void prio_changed_fair (struct rq *rq,struct task_struct *p,int oldprio,int running)
kernel/sched_fair.c static void task_new_fair(struct rq *rq, struct task_struct *p) · line 1701
Quoted 10 lines
kernel/sched_fair.c +9
-0{ +struct cfs_rq *cfs_rq =task_cfs_rq (p); +struct sched_entity *se = &p->se; +int on_rq = se->on_rq; + +if (p->prio < oldprio && on_rq) { +dequeue_entity (cfs_rq, se,0 ); + se->vruntime = cfs_rq->min_vruntime +sched_vslice (cfs_rq, se); +enqueue_entity (cfs_rq, se,0 ); + }
we very likely just enqueued the thing, and now we dequeue/enqueue again.. not very nice.
Quoted 3 lines
kernel/sched_fair.c +0
-0/* * Reschedule if we are currently running on this runqueue and * our priority decreased, or if we are not currently running on
Quoted 10 lines · Peter Zijlstra
Quoted 10 lines
kernel/sched_fair.c +4
-0@@ -1690,6 +1690,15 @@ static void task_new_fair(struct rq *rq, struct task_struct *p)static void prio_changed_fair (struct rq *rq,struct task_struct *p,int oldprio,int running) { +struct cfs_rq *cfs_rq =task_cfs_rq (p); +struct sched_entity *se = &p->se; +int on_rq = se->on_rq; +
Quoted 8 lines
kernel/sched_fair.c static void task_new_fair(struct rq *rq, struct task_struct *p) · line 1701
Quoted 5 lines
kernel/sched_fair.c +5
-0+if (p->prio < oldprio && on_rq) { +dequeue_entity (cfs_rq, se,0 ); + se->vruntime = cfs_rq->min_vruntime +sched_vslice (cfs_rq, se); +enqueue_entity (cfs_rq, se,0 ); + }we very likely just enqueued the thing, and now we dequeue/enqueue again.. not very nice.
but priority-change is a slowpath, so it shouldnt matter much, should it?
Ingo