Skip to contentSkip to conversation

nohz: A few improvements v2

PATCH In review Frederic Weisbecker 2 participants 12 messages Apr 21, 2015 – Apr 22, 2015

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover nohz: A few improvements v2

2nd version of the patchset "context_tracking: A few improvements" with the nohz isolcpus patches from Chris.

git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks.git
nohz/core

HEAD: b885e51f754a09bce2709c789a913735ee62bfd2

Thanks,
Frederic
---

Chris Metcalf (2):
nohz: Add tick_nohz_full_add_cpus_to() API
nohz: Set isolcpus when nohz_full is set

Frederic Weisbecker (2):
context_tracking: Protect against recursion
context_tracking: Inherit TIF_NOHZ through forks instead of context switches

 include/linux/context_tracking.h       | 10 -----
 include/linux/context_tracking_state.h |  1 +
 include/linux/tick.h                   |  7 ++++
 kernel/context_tracking.c              | 75 ++++++++++++++++++++++++----------
 kernel/sched/core.c                    |  4 +-
 5 files changed, 64 insertions(+), 33 deletions(-)
#1 lore raw

PATCH 1/4 context_tracking: Protect against recursion

Context tracking recursion can happen when an exception triggers in the middle of a call to a context tracking probe.

This special case can be caused by vmalloc faults. If an access to a memory area allocated by vmalloc happens in the middle of context_tracking_enter(), we may run into an endless fault loop because the exception in turn calls context_tracking_enter() which faults on the same vmalloc'ed memory, triggering an exception again, etc...

Some rare crashes have been reported so lets protect against this with a recursion counter.

Cc Thomas Gleixner <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Rik van Riel <[email protected]>
Cc Paul E. McKenney <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Dave Jones <[email protected]>
Cc Oleg Nesterov <[email protected]>
---
 include/linux/context_tracking_state.h |  1 +
 kernel/context_tracking.c              | 30 ++++++++++++++++++++++++++++++
 2 files changed, 31 insertions(+)
include/linux/context_tracking_state.h +1 -0
index 6b7b96a..678ecdf 100644
@@ -12,6 +12,7 @@ struct context_tracking {
 	 * may be further optimized using static keys.
 	 */
 	bool active;
+	int recursion;
 	enum ctx_state {
 		CONTEXT_KERNEL = 0,
 		CONTEXT_USER,
kernel/context_tracking.c +30 -0
index 72d59a1..b9e0b4f 100644
@@ -38,6 +38,25 @@ void context_tracking_cpu_set(int cpu)
 	}
 }
 
+static bool context_tracking_recursion_enter(void)
+{
+	int recursion;
+
+	recursion = __this_cpu_inc_return(context_tracking.recursion);
+	if (recursion == 1)
+		return true;
+
+	WARN_ONCE((recursion < 1), "Invalid context tracking recursion value %d\n", recursion);
+	__this_cpu_dec(context_tracking.recursion);
+
+	return false;
+}
+
+static void context_tracking_recursion_exit(void)
+{
+	__this_cpu_dec(context_tracking.recursion);
+}
+
 /**
  * context_tracking_enter - Inform the context tracking that the CPU is going
  *                          enter user or guest space mode.
@@ -75,6 +94,11 @@ void context_tracking_enter(enum ctx_state state)
 	WARN_ON_ONCE(!current->mm);
 
 	local_irq_save(flags);
+	if (!context_tracking_recursion_enter()) {
+		local_irq_restore(flags);
+		return;
+	}
+
 	if ( __this_cpu_read(context_tracking.state) != state) {
 		if (__this_cpu_read(context_tracking.active)) {
 			/*
@@ -105,6 +129,7 @@ void context_tracking_enter(enum ctx_state state)
 		 */
 		__this_cpu_write(context_tracking.state, state);
 	}
+	context_tracking_recursion_exit();
 	local_irq_restore(flags);
 }
 NOKPROBE_SYMBOL(context_tracking_enter);
@@ -139,6 +164,10 @@ void context_tracking_exit(enum ctx_state state)
 		return;
 
 	local_irq_save(flags);
+	if (!context_tracking_recursion_enter()) {
+		local_irq_restore(flags);
+		return;
+	}
 	if (__this_cpu_read(context_tracking.state) == state) {
 		if (__this_cpu_read(context_tracking.active)) {
 			/*
@@ -153,6 +182,7 @@ void context_tracking_exit(enum ctx_state state)
 		}
 		__this_cpu_write(context_tracking.state, CONTEXT_KERNEL);
 	}
+	context_tracking_recursion_exit();
 	local_irq_restore(flags);
 }
 NOKPROBE_SYMBOL(context_tracking_exit);
Signature · 1 line
2.1.4
#2 lore raw

PATCH 2/4 context_tracking: Inherit TIF_NOHZ through forks instead of context switches

In reply to Frederic Weisbecker

TIF_NOHZ is used by context_tracking to force syscall slow-path on every task in order to track userspace roundtrips. As such, it must be set on all running tasks.

It's currently explicitly inherited through context switches. There is no need to do it on this fast-path though. The flag could be simply set once for all on all tasks, whether they are running or not.

Lets do this by setting the flag to init task on early boot and let it propagate through fork inheritance.

While at it mark context_tracking_cpu_set() as init code, we only need it at early boot time.

Cc Thomas Gleixner <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Rik van Riel <[email protected]>
Cc Paul E. McKenney <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Dave Jones <[email protected]>
Cc Oleg Nesterov <[email protected]>
---
 include/linux/context_tracking.h | 10 --------
 kernel/context_tracking.c        | 51 ++++++++++++++++++++--------------------
 kernel/sched/core.c              |  1 -
 3 files changed, 26 insertions(+), 36 deletions(-)
include/linux/context_tracking.h +0 -10
index 2821838..b96bd29 100644
@@ -14,8 +14,6 @@ extern void context_tracking_enter(enum ctx_state state);
 extern void context_tracking_exit(enum ctx_state state);
 extern void context_tracking_user_enter(void);
 extern void context_tracking_user_exit(void);
-extern void __context_tracking_task_switch(struct task_struct *prev,
-					   struct task_struct *next);
 
 static inline void user_enter(void)
 {
@@ -51,19 +49,11 @@ static inline void exception_exit(enum ctx_state prev_ctx)
 	}
 }
 
-static inline void context_tracking_task_switch(struct task_struct *prev,
-						struct task_struct *next)
-{
-	if (context_tracking_is_enabled())
-		__context_tracking_task_switch(prev, next);
-}
 #else
 static inline void user_enter(void) { }
 static inline void user_exit(void) { }
 static inline enum ctx_state exception_enter(void) { return 0; }
 static inline void exception_exit(enum ctx_state prev_ctx) { }
-static inline void context_tracking_task_switch(struct task_struct *prev,
-						struct task_struct *next) { }
 #endif /* !CONFIG_CONTEXT_TRACKING */
 
 
kernel/context_tracking.c +26 -25

64 lines not on this page · raw message

kernel/sched/core.c +0 -1
index f9123a8..8414a6a 100644
@@ -2347,7 +2347,6 @@ context_switch(struct rq *rq, struct task_struct *prev,
 	 */
 	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
 
-	context_tracking_task_switch(prev, next);
 	/* Here we just switch the register state and the stack. */
 	switch_to(prev, next, prev);
 	barrier();
Signature · 1 line
2.1.4
maintainer
#3 lore raw

Re: [PATCH 2/4] context_tracking: Inherit TIF_NOHZ through forks instead of context switches

Quoted 14 lines · Frederic Weisbecker
kernel/context_tracking.c +13 -0
+void __init context_tracking_cpu_set(int cpu)
 {
+	static __initdata bool initialized = false;
+	struct task_struct *p, *t;
+
+	if (!per_cpu(context_tracking.active, cpu)) {
+		per_cpu(context_tracking.active, cpu) = true;
+		static_key_slow_inc(&context_tracking_enabled);
+	}
+
+	if (initialized)
+		return;
+
+	set_tsk_thread_flag(&init_task, TIF_NOHZ);

kernel/context_tracking.c void context_tracking_user_exit(void) · line 212

Quoted 10 lines
kernel/context_tracking.c +10 -0
+
+	/*
+	 * There shouldn't be any thread at this early boot stage
+	 * but the scheduler is ready to host any. So lets walk
+	 * the tasklist just in case. tasklist_lock isn't necessary
+	 * either that early but take it for correctness checkers.
+	 */
+	read_lock(&tasklist_lock);
+	for_each_process_thread(p, t)
+		set_tsk_thread_flag(t, TIF_NOHZ);

If there should not be any task, should there not be a WARN_ON_ONCE() here?

Quoted 4 lines
kernel/context_tracking.c +3 -0
+	read_unlock(&tasklist_lock);
+
+	initialized = true;
 }
#4 lore raw

Re: [PATCH 2/4] context_tracking: Inherit TIF_NOHZ through forks instead of context switches

Quoted 20 lines · Peter Zijlstra
Quoted 19 lines · Frederic Weisbecker
kernel/context_tracking.c +18 -0
+void __init context_tracking_cpu_set(int cpu)
 {
+	static __initdata bool initialized = false;
+	struct task_struct *p, *t;
+
+	if (!per_cpu(context_tracking.active, cpu)) {
+		per_cpu(context_tracking.active, cpu) = true;
+		static_key_slow_inc(&context_tracking_enabled);
+	}
+
+	if (initialized)
+		return;
+
+	set_tsk_thread_flag(&init_task, TIF_NOHZ);
+
+	/*
+	 * There shouldn't be any thread at this early boot stage
+	 * but the scheduler is ready to host any. So lets walk
+	 * the tasklist just in case. tasklist_lock isn't necessary
Quoted 8 lines

kernel/context_tracking.c void context_tracking_user_exit(void) · line 212

Quoted 5 lines
kernel/context_tracking.c +5 -0
+	 * either that early but take it for correctness checkers.
+	 */
+	read_lock(&tasklist_lock);
+	for_each_process_thread(p, t)
+		set_tsk_thread_flag(t, TIF_NOHZ);

If there should not be any task, should there not be a WARN_ON_ONCE() here?

Well, it's legal to have a task at that time because sched_init() was called. I just haven't observed any task other than init/0. But future code (or alternate configs than mine) might create a task between sched_init() and tick_init(). And the above code takes care of such a possibility.

Quoted 5 lines
Quoted 5 lines
kernel/context_tracking.c +3 -0
+	read_unlock(&tasklist_lock);
+
+	initialized = true;
 }
maintainer
#5 lore raw

Re: [PATCH 2/4] context_tracking: Inherit TIF_NOHZ through forks instead of context switches

Quoted 30 lines · Frederic Weisbecker
Quoted 20 lines · Peter Zijlstra
Quoted 19 lines · Frederic Weisbecker
kernel/context_tracking.c +18 -0
+void __init context_tracking_cpu_set(int cpu)
 {
+	static __initdata bool initialized = false;
+	struct task_struct *p, *t;
+
+	if (!per_cpu(context_tracking.active, cpu)) {
+		per_cpu(context_tracking.active, cpu) = true;
+		static_key_slow_inc(&context_tracking_enabled);
+	}
+
+	if (initialized)
+		return;
+
+	set_tsk_thread_flag(&init_task, TIF_NOHZ);
+
+	/*
+	 * There shouldn't be any thread at this early boot stage
+	 * but the scheduler is ready to host any. So lets walk
+	 * the tasklist just in case. tasklist_lock isn't necessary
Quoted 8 lines

kernel/context_tracking.c void context_tracking_user_exit(void) · line 212

Quoted 5 lines
kernel/context_tracking.c +5 -0
+	 * either that early but take it for correctness checkers.
+	 */
+	read_lock(&tasklist_lock);
+	for_each_process_thread(p, t)
+		set_tsk_thread_flag(t, TIF_NOHZ);

If there should not be any task, should there not be a WARN_ON_ONCE() here?

Quoted 4 lines

Well, it's legal to have a task at that time because sched_init() was called. I just haven't observed any task other than init/0. But future code (or alternate configs than mine) might create a task between sched_init() and tick_init(). And the above code takes care of such a possibility.

So why not do it sooner?

#6 lore raw

Re: [PATCH 2/4] context_tracking: Inherit TIF_NOHZ through forks instead of context switches

Quoted 31 lines · Peter Zijlstra
Quoted 30 lines · Frederic Weisbecker
Quoted 28 lines · Peter Zijlstra
Quoted 14 lines · Frederic Weisbecker
kernel/context_tracking.c +13 -0
+void __init context_tracking_cpu_set(int cpu)
 {
+	static __initdata bool initialized = false;
+	struct task_struct *p, *t;
+
+	if (!per_cpu(context_tracking.active, cpu)) {
+		per_cpu(context_tracking.active, cpu) = true;
+		static_key_slow_inc(&context_tracking_enabled);
+	}
+
+	if (initialized)
+		return;
+
+	set_tsk_thread_flag(&init_task, TIF_NOHZ);

kernel/context_tracking.c void context_tracking_user_exit(void) · line 212

Quoted 10 lines
kernel/context_tracking.c +10 -0
+
+	/*
+	 * There shouldn't be any thread at this early boot stage
+	 * but the scheduler is ready to host any. So lets walk
+	 * the tasklist just in case. tasklist_lock isn't necessary
+	 * either that early but take it for correctness checkers.
+	 */
+	read_lock(&tasklist_lock);
+	for_each_process_thread(p, t)
+		set_tsk_thread_flag(t, TIF_NOHZ);

If there should not be any task, should there not be a WARN_ON_ONCE() here?

Quoted 6 lines
Quoted 4 lines

Well, it's legal to have a task at that time because sched_init() was called. I just haven't observed any task other than init/0. But future code (or alternate configs than mine) might create a task between sched_init() and tick_init(). And the above code takes care of such a possibility.

So why not do it sooner?

Because I need the tick nohz intialization to be after irq initialization (init_IRQ()), which it depends on.

maintainer
#7 lore raw

Re: [PATCH 2/4] context_tracking: Inherit TIF_NOHZ through forks instead of context switches

Quoted 3 lines · Frederic Weisbecker

Because I need the tick nohz intialization to be after irq initialization (init_IRQ()), which it depends on.

OK, that's all _way_ before rest_init() which does the very first fork().

I think we can safely put a WARN in there.

#8 lore raw

Re: [PATCH 2/4] context_tracking: Inherit TIF_NOHZ through forks instead of context switches

Quoted 9 lines · Peter Zijlstra
Quoted 3 lines · Frederic Weisbecker

Because I need the tick nohz intialization to be after irq initialization (init_IRQ()), which it depends on.

OK, that's all _way_ before rest_init() which does the very first fork().

I think we can safely put a WARN in there.

Ah good, I'm doing that!

thanks!

#9 lore raw

PATCH 3/4 nohz: Add tick_nohz_full_add_cpus_to() API

In reply to Frederic Weisbecker

From: Chris Metcalf <[email protected]>

This API is useful to modify a cpumask indicating some special nohz-type functionality so that the nohz cores are automatically added to that set.

---
 include/linux/tick.h | 7 +++++++
 1 file changed, 7 insertions(+)
include/linux/tick.h +7 -0
index f8492da5..4191b56 100644
@@ -134,6 +134,12 @@ static inline bool tick_nohz_full_cpu(int cpu)
 	return cpumask_test_cpu(cpu, tick_nohz_full_mask);
 }
 
+static inline void tick_nohz_full_add_cpus_to(struct cpumask *mask)
+{
+	if (tick_nohz_full_enabled())
+		cpumask_or(mask, mask, tick_nohz_full_mask);
+}
+
 extern void __tick_nohz_full_check(void);
 extern void tick_nohz_full_kick(void);
 extern void tick_nohz_full_kick_cpu(int cpu);
@@ -142,6 +148,7 @@ extern void __tick_nohz_task_switch(struct task_struct *tsk);
 #else
 static inline bool tick_nohz_full_enabled(void) { return false; }
 static inline bool tick_nohz_full_cpu(int cpu) { return false; }
+static inline void tick_nohz_full_add_cpus_to(struct cpumask *mask) { }
 static inline void __tick_nohz_full_check(void) { }
 static inline void tick_nohz_full_kick_cpu(int cpu) { }
 static inline void tick_nohz_full_kick(void) { }
Signature · 1 line
2.1.4
#10 lore raw

PATCH 4/4 nohz: Set isolcpus when nohz_full is set

In reply to Frederic Weisbecker

From: Chris Metcalf <[email protected]>

nohz_full is only useful with isolcpus also set, since otherwise the scheduler has to run periodically to try to determine whether to steal work from other cores.

Accordingly, when booting with nohz_full=xxx on the command line, we should act as if isolcpus=xxx was also set, and set (or extend) the isolcpus set to include the nohz_full cpus.

---
 kernel/sched/core.c | 3 +++
 1 file changed, 3 insertions(+)
kernel/sched/core.c +3 -0
index 8414a6a..ee4f735 100644
@@ -7051,6 +7051,9 @@ void __init sched_init_smp(void)
 	alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
 	alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
 
+	/* nohz_full won't take effect without isolating the cpus. */
+	tick_nohz_full_add_cpus_to(cpu_isolated_map);
+
 	sched_init_numa();
 
 	/*
Signature · 1 line
2.1.4