Skip to contentSkip to conversation

sched/fair: sync expires_seq in distribute_cfs_runtime()

PATCH In review Cong Wang 3 participants 11 messages Jul 28, 2018 – Aug 3, 2018 Scheduler

lore
Share & tools
lore

Cong Wang

author
#0 lore raw

PATCH 1/1 sched/fair: sync expires_seq in distribute_cfs_runtime()

Each time we sync cfs_rq->runtime_expires with cfs_b->runtime_expires, we should sync its ->expires_seq too. However it is missing for distribute_cfs_runtime(), especially the slack timer call path.

Fixes 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition")
Cc Xunlei Pang <[email protected]>
Cc Ben Segall <[email protected]>
Cc Linus Torvalds <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
---
 kernel/sched/fair.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)
kernel/sched/fair.c +8 -4
index 2f0a0be4d344..910c50db3d74 100644
@@ -4857,7 +4857,7 @@ void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
 }
 
 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
-		u64 remaining, u64 expires)
+		u64 remaining, u64 expires, int expires_seq)
 {
 	struct cfs_rq *cfs_rq;
 	u64 runtime;
@@ -4880,6 +4880,7 @@ static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
 
 		cfs_rq->runtime_remaining += runtime;
 		cfs_rq->runtime_expires = expires;
+		cfs_rq->expires_seq = expires_seq;
 
 		/* we check whether we're throttled above */
 		if (cfs_rq->runtime_remaining > 0)
@@ -4905,7 +4906,7 @@ static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
 {
 	u64 runtime, runtime_expires;
-	int throttled;
+	int throttled, expires_seq;
 
 	/* no need to continue the timer with no bandwidth constraint */
 	if (cfs_b->quota == RUNTIME_INF)
@@ -4933,6 +4934,7 @@ static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
 	cfs_b->nr_throttled += overrun;
 
 	runtime_expires = cfs_b->runtime_expires;
+	expires_seq = cfs_b->expires_seq;
 
 	/*
 	 * This check is repeated as we are holding onto the new bandwidth while
@@ -4946,7 +4948,7 @@ static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
 		raw_spin_unlock(&cfs_b->lock);
 		/* we can't nest cfs_b->lock while distributing bandwidth */
 		runtime = distribute_cfs_runtime(cfs_b, runtime,
-						 runtime_expires);
+						 runtime_expires, expires_seq);
 		raw_spin_lock(&cfs_b->lock);
 
 		throttled = !list_empty(&cfs_b->throttled_cfs_rq);
@@ -5055,6 +5057,7 @@ static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
 {
 	u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
+	int expires_seq;
 	u64 expires;
 
 	/* confirm we're still not at a refresh boundary */
@@ -5068,12 +5071,13 @@ static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
 		runtime = cfs_b->runtime;
 
 	expires = cfs_b->runtime_expires;
+	expires_seq = cfs_b->expires_seq;
 	raw_spin_unlock(&cfs_b->lock);
 
 	if (!runtime)
 		return;
 
-	runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
+	runtime = distribute_cfs_runtime(cfs_b, runtime, expires, expires_seq);
 
 	raw_spin_lock(&cfs_b->lock);
 	if (expires == cfs_b->runtime_expires)
Signature · 1 line
2.14.4
#1 lore raw

Hi Cong,

Quoted 3 lines · Cong Wang

Each time we sync cfs_rq->runtime_expires with cfs_b->runtime_expires, we should sync its ->expires_seq too. However it is missing for distribute_cfs_runtime(), especially the slack timer call path.

I don't think it's a problem, as expires_seq will get synced in assign_cfs_rq_runtime().

Thanks,
Xunlei

Quoted 83 lines
Fixes 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition")
Cc Xunlei Pang <[email protected]>
Cc Ben Segall <[email protected]>
Cc Linus Torvalds <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
---
 kernel/sched/fair.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)
kernel/sched/fair.c +8 -4
index 2f0a0be4d344..910c50db3d74 100644
@@ -4857,7 +4857,7 @@ void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
 }
 
 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
-		u64 remaining, u64 expires)
+		u64 remaining, u64 expires, int expires_seq)
 {
 	struct cfs_rq *cfs_rq;
 	u64 runtime;
@@ -4880,6 +4880,7 @@ static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
 
 		cfs_rq->runtime_remaining += runtime;
 		cfs_rq->runtime_expires = expires;
+		cfs_rq->expires_seq = expires_seq;
 
 		/* we check whether we're throttled above */
 		if (cfs_rq->runtime_remaining > 0)
@@ -4905,7 +4906,7 @@ static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
 {
 	u64 runtime, runtime_expires;
-	int throttled;
+	int throttled, expires_seq;
 
 	/* no need to continue the timer with no bandwidth constraint */
 	if (cfs_b->quota == RUNTIME_INF)
@@ -4933,6 +4934,7 @@ static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
 	cfs_b->nr_throttled += overrun;
 
 	runtime_expires = cfs_b->runtime_expires;
+	expires_seq = cfs_b->expires_seq;
 
 	/*
 	 * This check is repeated as we are holding onto the new bandwidth while
@@ -4946,7 +4948,7 @@ static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
 		raw_spin_unlock(&cfs_b->lock);
 		/* we can't nest cfs_b->lock while distributing bandwidth */
 		runtime = distribute_cfs_runtime(cfs_b, runtime,
-						 runtime_expires);
+						 runtime_expires, expires_seq);
 		raw_spin_lock(&cfs_b->lock);
 
 		throttled = !list_empty(&cfs_b->throttled_cfs_rq);
@@ -5055,6 +5057,7 @@ static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
 {
 	u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
+	int expires_seq;
 	u64 expires;
 
 	/* confirm we're still not at a refresh boundary */
@@ -5068,12 +5071,13 @@ static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
 		runtime = cfs_b->runtime;
 
 	expires = cfs_b->runtime_expires;
+	expires_seq = cfs_b->expires_seq;
 	raw_spin_unlock(&cfs_b->lock);
 
 	if (!runtime)
 		return;
 
-	runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
+	runtime = distribute_cfs_runtime(cfs_b, runtime, expires, expires_seq);
 
 	raw_spin_lock(&cfs_b->lock);
 	if (expires == cfs_b->runtime_expires)
#2 lore raw
Quoted 7 lines · Xunlei Pang

Hi Cong,

Quoted 3 lines · Cong Wang

Each time we sync cfs_rq->runtime_expires with cfs_b->runtime_expires, we should sync its ->expires_seq too. However it is missing for distribute_cfs_runtime(), especially the slack timer call path.

Quoted 5 lines

I don't think it's a problem, as expires_seq will get synced in assign_cfs_rq_runtime().

Thanks,
Xunlei

It does seem unlikely to actually come up since the cfs_rq would have to not run until the period was expired-locally-but-not-globally, but there's no reason to not fix it.

Quoted 84 lines
Quoted 84 lines
Fixes 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition")
Cc Xunlei Pang <[email protected]>
Cc Ben Segall <[email protected]>
Cc Linus Torvalds <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
---
 kernel/sched/fair.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)
kernel/sched/fair.c +8 -4
index 2f0a0be4d344..910c50db3d74 100644
@@ -4857,7 +4857,7 @@ void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
 }
 
 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
-		u64 remaining, u64 expires)
+		u64 remaining, u64 expires, int expires_seq)
 {
 	struct cfs_rq *cfs_rq;
 	u64 runtime;
@@ -4880,6 +4880,7 @@ static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
 
 		cfs_rq->runtime_remaining += runtime;
 		cfs_rq->runtime_expires = expires;
+		cfs_rq->expires_seq = expires_seq;
 
 		/* we check whether we're throttled above */
 		if (cfs_rq->runtime_remaining > 0)
@@ -4905,7 +4906,7 @@ static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
 {
 	u64 runtime, runtime_expires;
-	int throttled;
+	int throttled, expires_seq;
 
 	/* no need to continue the timer with no bandwidth constraint */
 	if (cfs_b->quota == RUNTIME_INF)
@@ -4933,6 +4934,7 @@ static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
 	cfs_b->nr_throttled += overrun;
 
 	runtime_expires = cfs_b->runtime_expires;
+	expires_seq = cfs_b->expires_seq;
 
 	/*
 	 * This check is repeated as we are holding onto the new bandwidth while
@@ -4946,7 +4948,7 @@ static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
 		raw_spin_unlock(&cfs_b->lock);
 		/* we can't nest cfs_b->lock while distributing bandwidth */
 		runtime = distribute_cfs_runtime(cfs_b, runtime,
-						 runtime_expires);
+						 runtime_expires, expires_seq);
 		raw_spin_lock(&cfs_b->lock);
 
 		throttled = !list_empty(&cfs_b->throttled_cfs_rq);
@@ -5055,6 +5057,7 @@ static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
 {
 	u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
+	int expires_seq;
 	u64 expires;
 
 	/* confirm we're still not at a refresh boundary */
@@ -5068,12 +5071,13 @@ static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
 		runtime = cfs_b->runtime;
 
 	expires = cfs_b->runtime_expires;
+	expires_seq = cfs_b->expires_seq;
 	raw_spin_unlock(&cfs_b->lock);
 
 	if (!runtime)
 		return;
 
-	runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
+	runtime = distribute_cfs_runtime(cfs_b, runtime, expires, expires_seq);
 
 	raw_spin_lock(&cfs_b->lock);
 	if (expires == cfs_b->runtime_expires)
author
#3 lore raw
In reply to Xunlei Pang
Quoted 10 lines · Xunlei Pang

Hi Cong,

Quoted 3 lines · Cong Wang

Each time we sync cfs_rq->runtime_expires with cfs_b->runtime_expires, we should sync its ->expires_seq too. However it is missing for distribute_cfs_runtime(), especially the slack timer call path.

I don't think it's a problem, as expires_seq will get synced in assign_cfs_rq_runtime().

Sure, but there is a small window during which they are not synced. Why do you want to wait until the next assign_cfs_rq_runtime() when you already know runtime_expires is synced?

Also, expire_cfs_rq_runtime() is called before assign_cfs_rq_runtime() inside __account_cfs_rq_runtime(), which means the check of cfs_rq->expires_seq is not accurate for unthrottling case if the clock drift happens soon enough?

#4 lore raw
Quoted 12 lines · Cong Wang
Quoted 10 lines · Xunlei Pang

Hi Cong,

Quoted 3 lines · Cong Wang

Each time we sync cfs_rq->runtime_expires with cfs_b->runtime_expires, we should sync its ->expires_seq too. However it is missing for distribute_cfs_runtime(), especially the slack timer call path.

I don't think it's a problem, as expires_seq will get synced in assign_cfs_rq_runtime().

Quoted 9 lines

Sure, but there is a small window during which they are not synced. Why do you want to wait until the next assign_cfs_rq_runtime() when you already know runtime_expires is synced?

Also, expire_cfs_rq_runtime() is called before assign_cfs_rq_runtime() inside __account_cfs_rq_runtime(), which means the check of cfs_rq->expires_seq is not accurate for unthrottling case if the clock drift happens soon enough?

expire_cfs_rq_runtime():

    if (cfs_rq->expires_seq == cfs_b->expires_seq) {
        /* extend local deadline, drift is bounded above by 2 ticks */
        cfs_rq->runtime_expires += TICK_NSEC;
    } else {
        /* global deadline is ahead, expiration has passed */
        cfs_rq->runtime_remaining = 0;
    }

So if clock drift happens soon, then expires_seq decides the correct thing we should do: if cfs_b->expires_seq advanced, then clear the stale cfs_rq->runtime_remaining from the slack timer of the past period, then assign_cfs_rq_runtime() will refresh them afterwards, otherwise it is a real clock drift. I am still not getting where the race is?

#5 lore raw
Quoted 32 lines · Xunlei Pang
Quoted 12 lines · Cong Wang
Quoted 10 lines · Xunlei Pang

Hi Cong,

Quoted 3 lines · Cong Wang

Each time we sync cfs_rq->runtime_expires with cfs_b->runtime_expires, we should sync its ->expires_seq too. However it is missing for distribute_cfs_runtime(), especially the slack timer call path.

I don't think it's a problem, as expires_seq will get synced in assign_cfs_rq_runtime().

Quoted 9 lines

Sure, but there is a small window during which they are not synced. Why do you want to wait until the next assign_cfs_rq_runtime() when you already know runtime_expires is synced?

Also, expire_cfs_rq_runtime() is called before assign_cfs_rq_runtime() inside __account_cfs_rq_runtime(), which means the check of cfs_rq->expires_seq is not accurate for unthrottling case if the clock drift happens soon enough?

expire_cfs_rq_runtime():

    if (cfs_rq->expires_seq == cfs_b->expires_seq) {
        /* extend local deadline, drift is bounded above by 2 ticks */
        cfs_rq->runtime_expires += TICK_NSEC;
    } else {
        /* global deadline is ahead, expiration has passed */
        cfs_rq->runtime_remaining = 0;
    }
Quoted 5 lines

So if clock drift happens soon, then expires_seq decides the correct thing we should do: if cfs_b->expires_seq advanced, then clear the stale cfs_rq->runtime_remaining from the slack timer of the past period, then assign_cfs_rq_runtime() will refresh them afterwards, otherwise it is a real clock drift. I am still not getting where the race is?

Nothing /important/ goes wrong because distribute_cfs_runtime only fills runtime_remaining up to 1, not a real amount.

author
#6 lore raw
Quoted 35 lines · On Tue, Jul 31, 2018 at 10:13 AM <[email protected]> wrote:
Quoted 32 lines · Xunlei Pang
Quoted 12 lines · Cong Wang
Quoted 10 lines · Xunlei Pang

Hi Cong,

Quoted 3 lines · Cong Wang

Each time we sync cfs_rq->runtime_expires with cfs_b->runtime_expires, we should sync its ->expires_seq too. However it is missing for distribute_cfs_runtime(), especially the slack timer call path.

I don't think it's a problem, as expires_seq will get synced in assign_cfs_rq_runtime().

Quoted 9 lines

Sure, but there is a small window during which they are not synced. Why do you want to wait until the next assign_cfs_rq_runtime() when you already know runtime_expires is synced?

Also, expire_cfs_rq_runtime() is called before assign_cfs_rq_runtime() inside __account_cfs_rq_runtime(), which means the check of cfs_rq->expires_seq is not accurate for unthrottling case if the clock drift happens soon enough?

expire_cfs_rq_runtime():

    if (cfs_rq->expires_seq == cfs_b->expires_seq) {
        /* extend local deadline, drift is bounded above by 2 ticks */
        cfs_rq->runtime_expires += TICK_NSEC;
    } else {
        /* global deadline is ahead, expiration has passed */
        cfs_rq->runtime_remaining = 0;
    }
Quoted 5 lines
Quoted 5 lines

So if clock drift happens soon, then expires_seq decides the correct thing we should do: if cfs_b->expires_seq advanced, then clear the stale cfs_rq->runtime_remaining from the slack timer of the past period, then assign_cfs_rq_runtime() will refresh them afterwards, otherwise it is a real clock drift. I am still not getting where the race is?

But expires_seq is supposed to be the same here, after distribute_cfs_runtime(), therefore runtime_remaining is not supposed to be cleared.

Which part do I misunderstand? expires_seq should not be same here? Or you are saying a wrongly clear of runtime_remaning is fine?

Quoted 3 lines

Nothing /important/ goes wrong because distribute_cfs_runtime only fills runtime_remaining up to 1, not a real amount.

No, runtime_remaining is updated right before expire_cfs_rq_runtime():

static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)

{
        /* dock delta_exec before expiring quota (as it could span periods) */
        cfs_rq->runtime_remaining -= delta_exec;
        expire_cfs_rq_runtime(cfs_rq);

so almost certainly it can't be 1.

Which means the following check could be passed:

4655 if (cfs_rq->runtime_remaining < 0)
4656 return;

therefore we are reaching the clock drift logic code inside expire_cfs_rq_runtime()
where expires_seq is supposed to be same as they should be sync'ed. Therefore without patch, we wrongly clear the runtime_remainng?

Thanks.

#7 lore raw
Quoted 42 lines · Cong Wang
Quoted 35 lines · On Tue, Jul 31, 2018 at 10:13 AM <[email protected]> wrote:
Quoted 32 lines · Xunlei Pang
Quoted 12 lines · Cong Wang
Quoted 10 lines · Xunlei Pang

Hi Cong,

Quoted 3 lines · Cong Wang

Each time we sync cfs_rq->runtime_expires with cfs_b->runtime_expires, we should sync its ->expires_seq too. However it is missing for distribute_cfs_runtime(), especially the slack timer call path.

I don't think it's a problem, as expires_seq will get synced in assign_cfs_rq_runtime().

Quoted 9 lines

Sure, but there is a small window during which they are not synced. Why do you want to wait until the next assign_cfs_rq_runtime() when you already know runtime_expires is synced?

Also, expire_cfs_rq_runtime() is called before assign_cfs_rq_runtime() inside __account_cfs_rq_runtime(), which means the check of cfs_rq->expires_seq is not accurate for unthrottling case if the clock drift happens soon enough?

expire_cfs_rq_runtime():

    if (cfs_rq->expires_seq == cfs_b->expires_seq) {
        /* extend local deadline, drift is bounded above by 2 ticks */
        cfs_rq->runtime_expires += TICK_NSEC;
    } else {
        /* global deadline is ahead, expiration has passed */
        cfs_rq->runtime_remaining = 0;
    }
Quoted 5 lines
Quoted 5 lines

So if clock drift happens soon, then expires_seq decides the correct thing we should do: if cfs_b->expires_seq advanced, then clear the stale cfs_rq->runtime_remaining from the slack timer of the past period, then assign_cfs_rq_runtime() will refresh them afterwards, otherwise it is a real clock drift. I am still not getting where the race is?

Quoted 7 lines

But expires_seq is supposed to be the same here, after distribute_cfs_runtime(), therefore runtime_remaining is not supposed to be cleared.

Which part do I misunderstand? expires_seq should not be same here? Or you are saying a wrongly clear of runtime_remaning is fine?

Let's see the unthrottle cases.
1. for the periodic timer
distribute_cfs_runtime updates the throttled cfs_rq->runtime_expires to be a new value, so expire_cfs_rq_runtime does nothing because of:
rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires < 0

Afterwards assign_cfs_rq_runtime() will sync its expires_seq.

2. for the slack timer
the two expires_seq should be the same, so if clock drift happens soon, expire_cfs_rq_runtime regards it as true clock drift:
cfs_rq->runtime_expires += TICK_NSEC
If it happens that global expires_seq advances, it also doesn't matter, expire_cfs_rq_runtime will clear the stale expire_cfs_rq_runtime as expected.

Quoted 7 lines
Quoted 4 lines

Nothing /important/ goes wrong because distribute_cfs_runtime only fills runtime_remaining up to 1, not a real amount.

No, runtime_remaining is updated right before expire_cfs_rq_runtime():

Quoted 7 lines

static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)

{
        /* dock delta_exec before expiring quota (as it could span periods) */
        cfs_rq->runtime_remaining -= delta_exec;
        expire_cfs_rq_runtime(cfs_rq);

so almost certainly it can't be 1.

I think Ben means it firstly gets a distributtion of 1 to run after unthrottling, soon it will have a negative runtime_remaining, and go to assign_cfs_rq_runtime().

Thanks,
Xunlei

Quoted 13 lines

Which means the following check could be passed:

4655 if (cfs_rq->runtime_remaining < 0)
4656 return;

therefore we are reaching the clock drift logic code inside expire_cfs_rq_runtime()
where expires_seq is supposed to be same as they should be sync'ed. Therefore without patch, we wrongly clear the runtime_remainng?

Thanks.

author
#8 lore raw
Quoted 8 lines · Xunlei Pang

Let's see the unthrottle cases.
1. for the periodic timer
distribute_cfs_runtime updates the throttled cfs_rq->runtime_expires to be a new value, so expire_cfs_rq_runtime does nothing because of:
rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires < 0

Afterwards assign_cfs_rq_runtime() will sync its expires_seq.

Is there any guarantee rq_clock(cfs_rq) is always ahead of cfs_rq->runtime_expires in this case?

I doubt, because cfs_rq->runtime_expires could be assigned by a sched_clock() on a different CPU running the periodic timer.

Also, rq_clock() is behind sched_clock() on the same CPU too, sometimes it is merely hundreds of nanoseconds, sometimes it is tens of thousands nanoseconds in my environment. (I have a different patch to address this, but still not sure if it is correct.)

Quoted 8 lines

2. for the slack timer
the two expires_seq should be the same, so if clock drift happens soon, expire_cfs_rq_runtime regards it as true clock drift:
cfs_rq->runtime_expires += TICK_NSEC
If it happens that global expires_seq advances, it also doesn't matter, expire_cfs_rq_runtime will clear the stale expire_cfs_rq_runtime as expected.

Hmm, looks like due to the runtime_refresh_within() check in slack timer.

Quoted 14 lines
Quoted 14 lines
Quoted 5 lines

Nothing /important/ goes wrong because distribute_cfs_runtime only fills runtime_remaining up to 1, not a real amount.

No, runtime_remaining is updated right before expire_cfs_rq_runtime():

static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)

{
        /* dock delta_exec before expiring quota (as it could span periods) */
        cfs_rq->runtime_remaining -= delta_exec;
        expire_cfs_rq_runtime(cfs_rq);
Quoted 5 lines
Quoted 1 line

so almost certainly it can't be 1.

I think Ben means it firstly gets a distributtion of 1 to run after unthrottling, soon it will have a negative runtime_remaining, and go to assign_cfs_rq_runtime().

That is obvious, being 1 in distribute_cfs_runtime is not relevant to the discussion here.

#9 lore raw
In reply to Cong Wang
Quoted 56 lines · Cong Wang
Quoted 35 lines · On Tue, Jul 31, 2018 at 10:13 AM <[email protected]> wrote:
Quoted 32 lines · Xunlei Pang
Quoted 12 lines · Cong Wang
Quoted 10 lines · Xunlei Pang

Hi Cong,

Quoted 3 lines · Cong Wang

Each time we sync cfs_rq->runtime_expires with cfs_b->runtime_expires, we should sync its ->expires_seq too. However it is missing for distribute_cfs_runtime(), especially the slack timer call path.

I don't think it's a problem, as expires_seq will get synced in assign_cfs_rq_runtime().

Quoted 9 lines

Sure, but there is a small window during which they are not synced. Why do you want to wait until the next assign_cfs_rq_runtime() when you already know runtime_expires is synced?

Also, expire_cfs_rq_runtime() is called before assign_cfs_rq_runtime() inside __account_cfs_rq_runtime(), which means the check of cfs_rq->expires_seq is not accurate for unthrottling case if the clock drift happens soon enough?

expire_cfs_rq_runtime():

    if (cfs_rq->expires_seq == cfs_b->expires_seq) {
        /* extend local deadline, drift is bounded above by 2 ticks */
        cfs_rq->runtime_expires += TICK_NSEC;
    } else {
        /* global deadline is ahead, expiration has passed */
        cfs_rq->runtime_remaining = 0;
    }
Quoted 5 lines
Quoted 5 lines

So if clock drift happens soon, then expires_seq decides the correct thing we should do: if cfs_b->expires_seq advanced, then clear the stale cfs_rq->runtime_remaining from the slack timer of the past period, then assign_cfs_rq_runtime() will refresh them afterwards, otherwise it is a real clock drift. I am still not getting where the race is?

But expires_seq is supposed to be the same here, after distribute_cfs_runtime(), therefore runtime_remaining is not supposed to be cleared.

Which part do I misunderstand? expires_seq should not be same here? Or you are saying a wrongly clear of runtime_remaning is fine?

Quoted 3 lines

Nothing /important/ goes wrong because distribute_cfs_runtime only fills runtime_remaining up to 1, not a real amount.

No, runtime_remaining is updated right before expire_cfs_rq_runtime():

Quoted 7 lines

static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)

{
        /* dock delta_exec before expiring quota (as it could span periods) */
        cfs_rq->runtime_remaining -= delta_exec;
        expire_cfs_rq_runtime(cfs_rq);

so almost certainly it can't be 1.

Yes, in practice what's actually going to happen is that the runtime_remaining will be put to 1 by distribute, the cfs_rq will be unthrottled, and then when it runs it will go negative immediately and hit the negative check in expires, so expires_seq being wrong will not actually matter. In addition, the worst thing that will happen if one of the account_cfs_rq_runtime(cfs_rq, 0) paths is hit first is that it will lose 1ns of quota, which also doesn't really matter.

author
#10 lore raw
Quoted 7 lines · On Wed, Aug 1, 2018 at 10:17 AM <[email protected]> wrote:

Yes, in practice what's actually going to happen is that the runtime_remaining will be put to 1 by distribute, the cfs_rq will be unthrottled, and then when it runs it will go negative immediately and hit the negative check in expires, so expires_seq being wrong will not actually matter. In addition, the worst thing that will happen if one of the account_cfs_rq_runtime(cfs_rq, 0) paths is hit first is that it will lose 1ns of quota, which also doesn't really matter.

Ah, I see.

Thanks!