Skip to contentSkip to conversation

sched: Fix missing preemption opportunity

GIT PULL Frederic Weisbecker 3 participants 4 messages Jan 22, 2015 – Feb 1, 2015 Scheduler

lore
Share & tools
lore
#0 lore raw

Ingo,

Please pull the sched/urgent branch that can be found at:

git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks.git
sched/urgent

HEAD: 4ccbe99852951957419bc616f888e297a478e50b

It's part of the preempt/schedule cleanups series suggested by Linus. I'm reworking these but this one commit is an exception because it's a bugfix. So I'm sending it now. I believe it's not a regression so it's perhaps too late for -rc5. I let you judge. The commit is based on -rc5 so it can be merged on sched/core otherwise.

Thanks,
Frederic
---

Frederic Weisbecker (1):
sched: Fix missing preemption check in cond_resched()

 kernel/sched/core.c | 40 +++++++++++++++++++---------------------
 1 file changed, 19 insertions(+), 21 deletions(-)

---
commit 4ccbe99852951957419bc616f888e297a478e50b

Author: Frederic Weisbecker <[email protected]>
Date:   Sun Dec 14 22:04:47 2014 +0100

sched: Fix missing preemption check in cond_resched()

If an interrupt fires in cond_resched(), between the call to __schedule() and the PREEMPT_ACTIVE count decrementation, and that interrupt sets TIF_NEED_RESCHED, the call to preempt_schedule_irq() will be ignored due to the PREEMPT_ACTIVE count. This kind of scenario, with irq preemption being delayed because it's interrupting a preempt-disabled area, is usually fixed up after preemption is re-enabled back with an explicit call to preempt_schedule().

This is what preempt_enable() does but a raw preempt count decrement as performed by __preempt_count_sub(PREEMPT_ACTIVE) doesn't handle delayed preemption check. Therefore when such a race happens, the rescheduling is going to be delayed until the next scheduler or preemption entrypoint. This can be a problem for scheduler latency sensitive workloads.

Lets fix that by consolidating cond_resched() with preempt_schedule() internals.

kernel/sched/core.c +19 -21
index c0accc0..c7ed25d 100644
@@ -2877,6 +2877,21 @@ void __sched schedule_preempt_disabled(void)
 	preempt_disable();
 }
 
+static void preempt_schedule_common(void)
+{
+	do {
+		__preempt_count_add(PREEMPT_ACTIVE);
+		__schedule();
+		__preempt_count_sub(PREEMPT_ACTIVE);
+
+		/*
+		 * Check again in case we missed a preemption opportunity
+		 * between schedule and now.
+		 */
+		barrier();
+	} while (need_resched());
+}
+
 #ifdef CONFIG_PREEMPT
 /*
  * this is the entry point to schedule() from in-kernel preemption
@@ -2892,17 +2907,7 @@ asmlinkage __visible void __sched notrace preempt_schedule(void)
 	if (likely(!preemptible()))
 		return;
 
-	do {
-		__preempt_count_add(PREEMPT_ACTIVE);
-		__schedule();
-		__preempt_count_sub(PREEMPT_ACTIVE);
-
-		/*
-		 * Check again in case we missed a preemption opportunity
-		 * between schedule and now.
-		 */
-		barrier();
-	} while (need_resched());
+	preempt_schedule_common();
 }
 NOKPROBE_SYMBOL(preempt_schedule);
 EXPORT_SYMBOL(preempt_schedule);
@@ -4202,17 +4207,10 @@ SYSCALL_DEFINE0(sched_yield)
 	return 0;
 }
 
-static void __cond_resched(void)
-{
-	__preempt_count_add(PREEMPT_ACTIVE);
-	__schedule();
-	__preempt_count_sub(PREEMPT_ACTIVE);
-}
-
 int __sched _cond_resched(void)
 {
 	if (should_resched()) {
-		__cond_resched();
+		preempt_schedule_common();
 		return 1;
 	}
 	return 0;
@@ -4237,7 +4235,7 @@ int __cond_resched_lock(spinlock_t *lock)
 	if (spin_needbreak(lock) || resched) {
 		spin_unlock(lock);
 		if (resched)
-			__cond_resched();
+			preempt_schedule_common();
 		else
 			cpu_relax();
 		ret = 1;
@@ -4253,7 +4251,7 @@ int __sched __cond_resched_softirq(void)
 
 	if (should_resched()) {
 		local_bh_enable();
-		__cond_resched();
+		preempt_schedule_common();
 		local_bh_disable();
 		return 1;
 	}
maintainer
#1 lore raw

I picked up the patch; will drop it if Ingo also does ;-)

Quoted 7 lines · Frederic Weisbecker

+++ b/kernel/sched/core.c

kernel/sched/core.c +2 -0
@@ -2877,6 +2877,21 @@ void __sched schedule_preempt_disabled(void)
 	preempt_disable();
 }
 
+static void preempt_schedule_common(void)
+{

kernel/sched/core.c void __sched schedule_preempt_disabled(void) · line 2891

Quoted 10 lines
kernel/sched/core.c +10 -0
+	do {
+		__preempt_count_add(PREEMPT_ACTIVE);
+		__schedule();
+		__preempt_count_sub(PREEMPT_ACTIVE);
+
+		/*
+		 * Check again in case we missed a preemption opportunity
+		 * between schedule and now.
+		 */
+		barrier();

I do however wonder about this barrier() here; why do we think we need it?

Is that because test_bit() it 'broken'? The bitops are typically atomic ops and atomic reads should be through a volatile cast (x86 constant_test_bit doesn't seem to do this).

Should we go audit and fix that?

Quoted 2 lines
kernel/sched/core.c +2 -0
+	} while (need_resched());
+}
#2 lore raw
Quoted 25 lines · Peter Zijlstra

I picked up the patch; will drop it if Ingo also does ;-)

Quoted 7 lines · Frederic Weisbecker

+++ b/kernel/sched/core.c

kernel/sched/core.c +2 -0
@@ -2877,6 +2877,21 @@ void __sched schedule_preempt_disabled(void)
 	preempt_disable();
 }
 
+static void preempt_schedule_common(void)
+{

kernel/sched/core.c void __sched schedule_preempt_disabled(void) · line 2891

Quoted 10 lines
kernel/sched/core.c +10 -0
+	do {
+		__preempt_count_add(PREEMPT_ACTIVE);
+		__schedule();
+		__preempt_count_sub(PREEMPT_ACTIVE);
+
+		/*
+		 * Check again in case we missed a preemption opportunity
+		 * between schedule and now.
+		 */
+		barrier();

I do however wonder about this barrier() here; why do we think we need it?

Quoted 5 lines

Is that because test_bit() it 'broken'? The bitops are typically atomic ops and atomic reads should be through a volatile cast (x86 constant_test_bit doesn't seem to do this).

Should we go audit and fix that?

I looked up with git blame and this was already there prior the first git commit v2.6.12 without appropriate explanation.

We must make sure that the PREEMPT_ACTIVE decrement is visible before we do the NEED_RESCHED test or an interrupt could spuriously miss a preempt_schedule_irq() opportunity.

__preempt_count_sub() in asm-generic is an inline, so an implicit barrier(). Only x86 overwrites it yet and it does so through an inline as well.

And __preempt_count_ops() must really imply a barrier() anyway, anything else would be insane (probably we should specify that in a comment somewhere). Although I see that responsability is taken from non-underscored callers...

#3 lore raw
In reply to Frederic Weisbecker
sched: Fix missing preemption opportunity
Commit-ID:  a18b5d01819235629289212ad428a5ee2b40f0d9
Gitweb:     http://git.kernel.org/tip/a18b5d01819235629289212ad428a5ee2b40f0d9
Author:     Frederic Weisbecker <[email protected]>
AuthorDate: Thu, 22 Jan 2015 18:08:04 +0100
Committer:  Ingo Molnar <[email protected]>
CommitDate: Fri, 30 Jan 2015 19:38:51 +0100

sched: Fix missing preemption opportunity

If an interrupt fires in cond_resched(), between the call to __schedule() and the PREEMPT_ACTIVE count decrementation, and that interrupt sets TIF_NEED_RESCHED, the call to preempt_schedule_irq() will be ignored due to the PREEMPT_ACTIVE count. This kind of scenario, with irq preemption being delayed because it's interrupting a preempt-disabled area, is usually fixed up after preemption is re-enabled back with an explicit call to preempt_schedule().

This is what preempt_enable() does but a raw preempt count decrement as performed by __preempt_count_sub(PREEMPT_ACTIVE) doesn't handle delayed preemption check. Therefore when such a race happens, the rescheduling is going to be delayed until the next scheduler or preemption entrypoint. This can be a problem for scheduler latency sensitive workloads.

Lets fix that by consolidating cond_resched() with preempt_schedule() internals.

---
 kernel/sched/core.c | 40 +++++++++++++++++++---------------------
 1 file changed, 19 insertions(+), 21 deletions(-)
kernel/sched/core.c +19 -21
index 0b591fe..54dce01 100644
@@ -2884,6 +2884,21 @@ void __sched schedule_preempt_disabled(void)
 	preempt_disable();
 }
 
+static void preempt_schedule_common(void)
+{
+	do {
+		__preempt_count_add(PREEMPT_ACTIVE);
+		__schedule();
+		__preempt_count_sub(PREEMPT_ACTIVE);
+
+		/*
+		 * Check again in case we missed a preemption opportunity
+		 * between schedule and now.
+		 */
+		barrier();
+	} while (need_resched());
+}
+
 #ifdef CONFIG_PREEMPT
 /*
  * this is the entry point to schedule() from in-kernel preemption
@@ -2899,17 +2914,7 @@ asmlinkage __visible void __sched notrace preempt_schedule(void)
 	if (likely(!preemptible()))
 		return;
 
-	do {
-		__preempt_count_add(PREEMPT_ACTIVE);
-		__schedule();
-		__preempt_count_sub(PREEMPT_ACTIVE);
-
-		/*
-		 * Check again in case we missed a preemption opportunity
-		 * between schedule and now.
-		 */
-		barrier();
-	} while (need_resched());
+	preempt_schedule_common();
 }
 NOKPROBE_SYMBOL(preempt_schedule);
 EXPORT_SYMBOL(preempt_schedule);
@@ -4209,17 +4214,10 @@ SYSCALL_DEFINE0(sched_yield)
 	return 0;
 }
 
-static void __cond_resched(void)
-{
-	__preempt_count_add(PREEMPT_ACTIVE);
-	__schedule();
-	__preempt_count_sub(PREEMPT_ACTIVE);
-}
-
 int __sched _cond_resched(void)
 {
 	if (should_resched()) {
-		__cond_resched();
+		preempt_schedule_common();
 		return 1;
 	}
 	return 0;
@@ -4244,7 +4242,7 @@ int __cond_resched_lock(spinlock_t *lock)
 	if (spin_needbreak(lock) || resched) {
 		spin_unlock(lock);
 		if (resched)
-			__cond_resched();
+			preempt_schedule_common();
 		else
 			cpu_relax();
 		ret = 1;
@@ -4260,7 +4258,7 @@ int __sched __cond_resched_softirq(void)
 
 	if (should_resched()) {
 		local_bh_enable();
-		__cond_resched();
+		preempt_schedule_common();
 		local_bh_disable();
 		return 1;
 	}