Skip to contentSkip to conversation

cpu_stop: implement and use cpu_stop, take#2

PATCH In review Tejun Heo 3 participants 15 messages May 4, 2010 – May 7, 2010

Subsystems (5)
lore
Share & tools
lore

Tejun Heo

author
#0 lore raw

PATCH 1/1 cpu_stop: implement and use cpu_stop, take#2

Hello,

This is the second take of cpu_stop. There have been only two minor changes from the last take[L].

* work_buf initialization in stop_one_cpu_nowait() simplified.

* Comment to clarify active_balance_work synchronization added.

Peter, if you ack the series, I'll add refresh the patches w/ your ACK and request pull into sched/core to Ingo.

This patchset contains the following four patches.

0001-cpu_stop-implement-stop_cpu-s.patch 0002-stop_machine-reimplement-using-cpu_stop.patch 0003-scheduler-replace-migration_thread-with-cpu_stop.patch 0004-scheduler-kill-paranoia-check-in-synchronize_sched_e.patch

The patches are against the current linux-2.6-tip/sched/core (99bd5e2f245d8cd17d040c82d40becdb3efd9b69) and are available in the following git tree.

git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git cpu_stop

diffstat follows.

 Documentation/RCU/torture.txt |   10 
 arch/s390/kernel/time.c       |    1 
 drivers/xen/manage.c          |   14 -
 include/linux/rcutiny.h       |    2 
 include/linux/rcutree.h       |    1 
 include/linux/stop_machine.h  |   59 ++--
 kernel/cpu.c                  |    8 
 kernel/module.c               |   14 -
 kernel/rcutorture.c           |    2 
 kernel/sched.c                |  271 +++------------------
 kernel/sched_fair.c           |   48 ++-
 kernel/stop_machine.c         |  530 ++++++++++++++++++++++++++++++++----------
 12 files changed, 525 insertions(+), 435 deletions(-)

Thanks.

--
tejun

[L] http://thread.gmane.org/gmane.linux.kernel/976691

author
#1 lore raw

PATCH 1/4 cpu_stop: implement stop_cpu[s]()

Implement a simplistic per-cpu maximum priority cpu monopolization mechanism. A non-sleeping callback can be scheduled to run on one or multiple cpus with maximum priority monopolozing those cpus. This is primarily to replace and unify RT workqueue usage in stop_machine and scheduler migration_thread which currently is serving multiple purposes.

Four functions are provided - stop_one_cpu(), stop_one_cpu_nowait(), stop_cpus() and try_stop_cpus().

This is to allow clean sharing of resources among stop_cpu and all the migration thread users. One stopper thread per cpu is created which is currently named "stopper/CPU". This will eventually replace the migration thread and take on its name.

* This facility was originally named cpuhog and lived in separate files but Peter Zijlstra nacked the name and thus got renamed to cpu_stop and moved into stop_machine.c.

* Better reporting of preemption leak as per Peter's suggestion.

Cc Oleg Nesterov <[email protected]>
Cc Dimitri Sivanich <[email protected]>
Cc Peter Zijlstra <[email protected]>
---
 include/linux/stop_machine.h |   39 ++++-
 kernel/stop_machine.c        |  372 +++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 402 insertions(+), 9 deletions(-)
include/linux/stop_machine.h +35 -4
index baba3a2..efcbd6c 100644
@@ -1,15 +1,46 @@
 #ifndef _LINUX_STOP_MACHINE
 #define _LINUX_STOP_MACHINE
-/* "Bogolock": stop the entire machine, disable interrupts.  This is a
-   very heavy lock, which is equivalent to grabbing every spinlock
-   (and more).  So the "read" side to such a lock is anything which
-   disables preeempt. */
+
 #include <linux/cpu.h>
 #include <linux/cpumask.h>
+#include <linux/list.h>
 #include <asm/system.h>
 
 #if defined(CONFIG_STOP_MACHINE) && defined(CONFIG_SMP)
 
+/*
+ * stop_cpu[s]() is simplistic per-cpu maximum priority cpu
+ * monopolization mechanism.  The caller can specify a non-sleeping
+ * function to be executed on a single or multiple cpus preempting all
+ * other processes and monopolizing those cpus until it finishes.
+ *
+ * Resources for this mechanism are preallocated when a cpu is brought
+ * up and requests are guaranteed to be served as long as the target
+ * cpus are online.
+ */
+
+typedef int (*cpu_stop_fn_t)(void *arg);
+
+struct cpu_stop_work {
+	struct list_head	list;		/* cpu_stopper->works */
+	cpu_stop_fn_t		fn;
+	void			*arg;
+	struct cpu_stop_done	*done;
+};
+
+int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg);
+void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
+			 struct cpu_stop_work *work_buf);
+int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg);
+int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg);
+
+/*
+ * stop_machine "Bogolock": stop the entire machine, disable
+ * interrupts.  This is a very heavy lock, which is equivalent to
+ * grabbing every spinlock (and more).  So the "read" side to such a
+ * lock is anything which disables preeempt.
+ */
+
 /**
  * stop_machine: freeze the machine on all CPUs and run this function
  * @fn: the function to run
kernel/stop_machine.c +367 -5

384 lines not on this page · raw message

Signature · 1 line
1.6.4.2
author
#2 lore raw

PATCH 2/4 stop_machine: reimplement using cpu_stop

In reply to Tejun Heo

Reimplement stop_machine using cpu_stop. As cpu stoppers are guaranteed to be available for all online cpus, stop_machine_create/destroy() are no longer necessary and removed.

With resource management and synchronization handled by cpu_stop, the new implementation is much simpler. Asking the cpu_stop to execute the stop_cpu() state machine on all online cpus with cpu hotplug disabled is enough.

stop_machine itself doesn't need to manage any global resources anymore, so all per-instance information is rolled into struct stop_machine_data and the mutex and all static data variables are removed.

The previous implementation created and destroyed RT workqueues as necessary which made stop_machine() calls highly expensive on very large machines. According to Dimitri Sivanich, preventing the dynamic creation/destruction makes booting faster more than twice on very large machines. cpu_stop resources are preallocated for all online cpus and should have the same effect.

---
 arch/s390/kernel/time.c      |    1 -
 drivers/xen/manage.c         |   14 +---
 include/linux/stop_machine.h |   20 -----
 kernel/cpu.c                 |    8 --
 kernel/module.c              |   14 +---
 kernel/stop_machine.c        |  158 ++++++++++--------------------------------
 6 files changed, 42 insertions(+), 173 deletions(-)
arch/s390/kernel/time.c +0 -1
index fba6dec..03d9656 100644
@@ -390,7 +390,6 @@ static void __init time_init_wq(void)
 	if (time_sync_wq)
 		return;
 	time_sync_wq = create_singlethread_workqueue("timesync");
-	stop_machine_create();
 }
 
 /*
drivers/xen/manage.c +2 -12
index 2ac4440..8943b8c 100644
@@ -80,12 +80,6 @@ static void do_suspend(void)
 
 	shutting_down = SHUTDOWN_SUSPEND;
 
-	err = stop_machine_create();
-	if (err) {
-		printk(KERN_ERR "xen suspend: failed to setup stop_machine %d\n", err);
-		goto out;
-	}
-
 #ifdef CONFIG_PREEMPT
 	/* If the kernel is preemptible, we need to freeze all the processes
 	   to prevent them from being in the middle of a pagetable update
@@ -93,7 +87,7 @@ static void do_suspend(void)
 	err = freeze_processes();
 	if (err) {
 		printk(KERN_ERR "xen suspend: freeze failed %d\n", err);
-		goto out_destroy_sm;
+		goto out;
 	}
 #endif
 
@@ -136,12 +130,8 @@ out_resume:
 out_thaw:
 #ifdef CONFIG_PREEMPT
 	thaw_processes();
-
-out_destroy_sm:
-#endif
-	stop_machine_destroy();
-
 out:
+#endif
 	shutting_down = SHUTDOWN_INVALID;
 }
 #endif	/* CONFIG_PM_SLEEP */
include/linux/stop_machine.h +0 -20
index efcbd6c..0e552e7 100644
@@ -67,23 +67,6 @@ int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus);
  */
 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus);
 
-/**
- * stop_machine_create: create all stop_machine threads
- *
- * Description: This causes all stop_machine threads to be created before
- * stop_machine actually gets called. This can be used by subsystems that
- * need a non failing stop_machine infrastructure.
- */
-int stop_machine_create(void);
-
-/**
- * stop_machine_destroy: destroy all stop_machine threads
- *
- * Description: This causes all stop_machine threads which were created with
- * stop_machine_create to be destroyed again.
- */
-void stop_machine_destroy(void);
-
 #else
 
 static inline int stop_machine(int (*fn)(void *), void *data,
@@ -96,8 +79,5 @@ static inline int stop_machine(int (*fn)(void *), void *data,
 	return ret;
 }
 
-static inline int stop_machine_create(void) { return 0; }
-static inline void stop_machine_destroy(void) { }
-
 #endif /* CONFIG_SMP */
 #endif /* _LINUX_STOP_MACHINE */
kernel/cpu.c +0 -8
index 914aedc..5457775 100644
@@ -266,9 +266,6 @@ int __ref cpu_down(unsigned int cpu)
 {
 	int err;
 
-	err = stop_machine_create();
-	if (err)
-		return err;
 	cpu_maps_update_begin();
 
 	if (cpu_hotplug_disabled) {
@@ -280,7 +277,6 @@ int __ref cpu_down(unsigned int cpu)
 
 out:
 	cpu_maps_update_done();
-	stop_machine_destroy();
 	return err;
 }
 EXPORT_SYMBOL(cpu_down);
@@ -361,9 +357,6 @@ int disable_nonboot_cpus(void)
 {
 	int cpu, first_cpu, error;
 
-	error = stop_machine_create();
-	if (error)
-		return error;
 	cpu_maps_update_begin();
 	first_cpu = cpumask_first(cpu_online_mask);
 	/*
@@ -394,7 +387,6 @@ int disable_nonboot_cpus(void)
 		printk(KERN_ERR "Non-boot CPUs are not disabled\n");
 	}
 	cpu_maps_update_done();
-	stop_machine_destroy();
 	return error;
 }
 
kernel/module.c +2 -12
index 1016b75..0838246 100644
@@ -723,16 +723,8 @@ SYSCALL_DEFINE2(delete_module, const char __user *, name_user,
 		return -EFAULT;
 	name[MODULE_NAME_LEN-1] = '\0';
 
-	/* Create stop_machine threads since free_module relies on
-	 * a non-failing stop_machine call. */
-	ret = stop_machine_create();
-	if (ret)
-		return ret;
-
-	if (mutex_lock_interruptible(&module_mutex) != 0) {
-		ret = -EINTR;
-		goto out_stop;
-	}
+	if (mutex_lock_interruptible(&module_mutex) != 0)
+		return -EINTR;
 
 	mod = find_module(name);
 	if (!mod) {
@@ -792,8 +784,6 @@ SYSCALL_DEFINE2(delete_module, const char __user *, name_user,
 
  out:
 	mutex_unlock(&module_mutex);
-out_stop:
-	stop_machine_destroy();
 	return ret;
 }
 
kernel/stop_machine.c +38 -120

212 lines not on this page · raw message

Signature · 1 line
1.6.4.2
author
#3 lore raw

PATCH 3/4 scheduler: replace migration_thread with cpu_stop

In reply to Tejun Heo

Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy.

This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary.

synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped.

synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree.

Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed.

This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes.

With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner.

Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Dipankar Sarma <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Paul E. McKenney <[email protected]>
Cc Oleg Nesterov <[email protected]>
Cc Dimitri Sivanich <[email protected]>
---
 Documentation/RCU/torture.txt |   10 --
 include/linux/rcutiny.h       |    2 -
 include/linux/rcutree.h       |    1 -
 kernel/rcutorture.c           |    2 +-
 kernel/sched.c                |  303 +++++++++++------------------------------
 kernel/sched_fair.c           |   48 +++++--
 kernel/stop_machine.c         |    2 +-
 7 files changed, 115 insertions(+), 253 deletions(-)
Documentation/RCU/torture.txt +0 -10
index 0e50bc2..5d90167 100644
@@ -182,16 +182,6 @@ Similarly, sched_expedited RCU provides the following:
 	sched_expedited-torture: Reader Pipe:  12660320201 95875 0 0 0 0 0 0 0 0 0
 	sched_expedited-torture: Reader Batch:  12660424885 0 0 0 0 0 0 0 0 0 0
 	sched_expedited-torture: Free-Block Circulation:  1090795 1090795 1090794 1090793 1090792 1090791 1090790 1090789 1090788 1090787 0
-	state: -1 / 0:0 3:0 4:0
-
-As before, the first four lines are similar to those for RCU.
-The last line shows the task-migration state.  The first number is
--1 if synchronize_sched_expedited() is idle, -2 if in the process of
-posting wakeups to the migration kthreads, and N when waiting on CPU N.
-Each of the colon-separated fields following the "/" is a CPU:state pair.
-Valid states are "0" for idle, "1" for waiting for quiescent state,
-"2" for passed through quiescent state, and "3" when a race with a
-CPU-hotplug event forces use of the synchronize_sched() primitive.
 
 
 USAGE
include/linux/rcutiny.h +0 -2
index a519587..0006b2d 100644
@@ -60,8 +60,6 @@ static inline long rcu_batches_completed_bh(void)
 	return 0;
 }
 
-extern int rcu_expedited_torture_stats(char *page);
-
 static inline void rcu_force_quiescent_state(void)
 {
 }
include/linux/rcutree.h +0 -1
index 42cc3a0..24e467e 100644
@@ -35,7 +35,6 @@ struct notifier_block;
 extern void rcu_sched_qs(int cpu);
 extern void rcu_bh_qs(int cpu);
 extern int rcu_needs_cpu(int cpu);
-extern int rcu_expedited_torture_stats(char *page);
 
 #ifdef CONFIG_TREE_PREEMPT_RCU
 
kernel/rcutorture.c +1 -1
index 58df55b..2b676f3 100644
@@ -669,7 +669,7 @@ static struct rcu_torture_ops sched_expedited_ops = {
 	.sync		= synchronize_sched_expedited,
 	.cb_barrier	= NULL,
 	.fqs		= rcu_sched_force_quiescent_state,
-	.stats		= rcu_expedited_torture_stats,
+	.stats		= NULL,
 	.irq_capable	= 1,
 	.name		= "sched_expedited"
 };
kernel/sched.c +79 -224

475 lines not on this page · raw message

kernel/sched_fair.c +34 -14

92 lines not on this page · raw message

kernel/stop_machine.c +1 -1
index 884c7a1..5b20141 100644
@@ -301,7 +301,7 @@ static int __cpuinit cpu_stop_cpu_callback(struct notifier_block *nfb,
 	case CPU_UP_PREPARE:
 		BUG_ON(stopper->thread || stopper->enabled ||
 		       !list_empty(&stopper->works));
-		p = kthread_create(cpu_stopper_thread, stopper, "stopper/%d",
+		p = kthread_create(cpu_stopper_thread, stopper, "migration/%d",
 				   cpu);
 		if (IS_ERR(p))
 			return NOTIFY_BAD;
Signature · 1 line
1.6.4.2
#4 lore raw

Re: [PATCH 3/4] scheduler: replace migration_thread with cpu_stop

Quoted 31 lines · Tejun Heo

Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy.

This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary.

synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped.

synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree.

Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed.

Quoted 7 lines

This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes.

With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner.

Nice!!!

I do have one concern, though. Can the following sequence of events occur?

o CPU 0 calls synchronize_sched_expedited().

o CPU 1 is already in cpu_stopper_thread(), perhaps being
interrupted or NMIed just after returning from cpu_stop_signal_done() in cpu_stopper_thread().

o Therefore, when CPU 0 queues the work for CPU 1, CPU 1
loops right around and processes it. There will be no context switch on CPU 1.

At first glance, this looks safe because:

1. Although there is no context switch, there (presumably)
can be no RCU read-side critical sections on CPU 1 that span this sequence of events. (As far as I can see, any such RCU read-side critical section would be due to abuse of rcu_read_lock() and friends.)

2. CPU 1 will acquire and release stopper->lock, and
further more will do an atomic_dec_and_test() in cpu_stop_signal_done(). The former is a weak guarantee, but the latter guarantees a memory barrier, so that any subsequent code on CPU 1 will be guaranteed to see changes on CPU 0 prior to the call to synchronize_sched_expedited().

The guarantee required is that there will be a full memory barrier on each affected CPU between the time that try_stop_cpus() is called and the time that it returns.

Both guarantees (1) and (2) are absolutely required for us to expect that synchronize_sched_expedited() will operate correctly. If my reading is correct, could you please document guarantee (2) in the docbook header for try_stop_cpus()? Otherwise, a fix is needed.

Or am I misunderstanding the code?

Thanx, Paul

Quoted 690 lines
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Dipankar Sarma <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Paul E. McKenney <[email protected]>
Cc Oleg Nesterov <[email protected]>
Cc Dimitri Sivanich <[email protected]>
---
 Documentation/RCU/torture.txt |   10 --
 include/linux/rcutiny.h       |    2 -
 include/linux/rcutree.h       |    1 -
 kernel/rcutorture.c           |    2 +-
 kernel/sched.c                |  303 +++++++++++------------------------------
 kernel/sched_fair.c           |   48 +++++--
 kernel/stop_machine.c         |    2 +-
 7 files changed, 115 insertions(+), 253 deletions(-)
Documentation/RCU/torture.txt +0 -10
index 0e50bc2..5d90167 100644
@@ -182,16 +182,6 @@ Similarly, sched_expedited RCU provides the following:
 	sched_expedited-torture: Reader Pipe:  12660320201 95875 0 0 0 0 0 0 0 0 0
 	sched_expedited-torture: Reader Batch:  12660424885 0 0 0 0 0 0 0 0 0 0
 	sched_expedited-torture: Free-Block Circulation:  1090795 1090795 1090794 1090793 1090792 1090791 1090790 1090789 1090788 1090787 0
-	state: -1 / 0:0 3:0 4:0
-
-As before, the first four lines are similar to those for RCU.
-The last line shows the task-migration state.  The first number is
--1 if synchronize_sched_expedited() is idle, -2 if in the process of
-posting wakeups to the migration kthreads, and N when waiting on CPU N.
-Each of the colon-separated fields following the "/" is a CPU:state pair.
-Valid states are "0" for idle, "1" for waiting for quiescent state,
-"2" for passed through quiescent state, and "3" when a race with a
-CPU-hotplug event forces use of the synchronize_sched() primitive.
 
 
 USAGE
include/linux/rcutiny.h +0 -2
index a519587..0006b2d 100644
@@ -60,8 +60,6 @@ static inline long rcu_batches_completed_bh(void)
 	return 0;
 }
 
-extern int rcu_expedited_torture_stats(char *page);
-
 static inline void rcu_force_quiescent_state(void)
 {
 }
include/linux/rcutree.h +0 -1
index 42cc3a0..24e467e 100644
@@ -35,7 +35,6 @@ struct notifier_block;
 extern void rcu_sched_qs(int cpu);
 extern void rcu_bh_qs(int cpu);
 extern int rcu_needs_cpu(int cpu);
-extern int rcu_expedited_torture_stats(char *page);
 
 #ifdef CONFIG_TREE_PREEMPT_RCU
 
kernel/rcutorture.c +1 -1
index 58df55b..2b676f3 100644
@@ -669,7 +669,7 @@ static struct rcu_torture_ops sched_expedited_ops = {
 	.sync		= synchronize_sched_expedited,
 	.cb_barrier	= NULL,
 	.fqs		= rcu_sched_force_quiescent_state,
-	.stats		= rcu_expedited_torture_stats,
+	.stats		= NULL,
 	.irq_capable	= 1,
 	.name		= "sched_expedited"
 };
kernel/sched.c +79 -224
index 4956ed0..8a198f1 100644
@@ -55,9 +55,9 @@
 #include <linux/cpu.h>
 #include <linux/cpuset.h>
 #include <linux/percpu.h>
-#include <linux/kthread.h>
 #include <linux/proc_fs.h>
 #include <linux/seq_file.h>
+#include <linux/stop_machine.h>
 #include <linux/sysctl.h>
 #include <linux/syscalls.h>
 #include <linux/times.h>
@@ -539,15 +539,13 @@ struct rq {
 	int post_schedule;
 	int active_balance;
 	int push_cpu;
+	struct cpu_stop_work active_balance_work;
 	/* cpu of this runqueue: */
 	int cpu;
 	int online;
 
 	unsigned long avg_load_per_task;
 
-	struct task_struct *migration_thread;
-	struct list_head migration_queue;
-
 	u64 rt_avg;
 	u64 age_stamp;
 	u64 idle_stamp;
@@ -2037,21 +2035,18 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
 	__set_task_cpu(p, new_cpu);
 }
 
-struct migration_req {
-	struct list_head list;
-
+struct migration_arg {
 	struct task_struct *task;
 	int dest_cpu;
-
-	struct completion done;
 };
 
+static int migration_cpu_stop(void *data);
+
 /*
  * The task's runqueue lock must be held.
  * Returns true if you have to wait for migration thread.
  */
-static int
-migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
+static bool migrate_task(struct task_struct *p, int dest_cpu)
 {
 	struct rq *rq = task_rq(p);
 
@@ -2059,15 +2054,7 @@ migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
 	 * If the task is not on a runqueue (and not running), then
 	 * the next wake-up will properly place the task.
 	 */
-	if (!p->se.on_rq && !task_running(rq, p))
-		return 0;
-
-	init_completion(&req->done);
-	req->task = p;
-	req->dest_cpu = dest_cpu;
-	list_add(&req->list, &rq->migration_queue);
-
-	return 1;
+	return p->se.on_rq || task_running(rq, p);
 }
 
 /*
@@ -3110,7 +3097,6 @@ static void update_cpu_load(struct rq *this_rq)
 void sched_exec(void)
 {
 	struct task_struct *p = current;
-	struct migration_req req;
 	unsigned long flags;
 	struct rq *rq;
 	int dest_cpu;
@@ -3124,17 +3110,11 @@ void sched_exec(void)
 	 * select_task_rq() can race against ->cpus_allowed
 	 */
 	if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
-	    likely(cpu_active(dest_cpu)) &&
-	    migrate_task(p, dest_cpu, &req)) {
-		/* Need to wait for migration thread (might exit: take ref). */
-		struct task_struct *mt = rq->migration_thread;
+	    likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
+		struct migration_arg arg = { p, dest_cpu };
 
-		get_task_struct(mt);
 		task_rq_unlock(rq, &flags);
-		wake_up_process(mt);
-		put_task_struct(mt);
-		wait_for_completion(&req.done);
-
+		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
 		return;
 	}
 unlock:
@@ -5290,17 +5270,15 @@ static inline void sched_init_granularity(void)
 /*
  * This is how migration works:
  *
- * 1) we queue a struct migration_req structure in the source CPU's
- *    runqueue and wake up that CPU's migration thread.
- * 2) we down() the locked semaphore => thread blocks.
- * 3) migration thread wakes up (implicitly it forces the migrated
- *    thread off the CPU)
- * 4) it gets the migration request and checks whether the migrated
- *    task is still in the wrong runqueue.
- * 5) if it's in the wrong runqueue then the migration thread removes
+ * 1) we invoke migration_cpu_stop() on the target CPU using
+ *    stop_one_cpu().
+ * 2) stopper starts to run (implicitly forcing the migrated thread
+ *    off the CPU)
+ * 3) it checks whether the migrated task is still in the wrong runqueue.
+ * 4) if it's in the wrong runqueue then the migration thread removes
  *    it and puts it into the right queue.
- * 6) migration thread up()s the semaphore.
- * 7) we wake up and the migration is done.
+ * 5) stopper completes and stop_one_cpu() returns and the migration
+ *    is done.
  */
 
 /*
@@ -5314,9 +5292,9 @@ static inline void sched_init_granularity(void)
  */
 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
 {
-	struct migration_req req;
 	unsigned long flags;
 	struct rq *rq;
+	unsigned int dest_cpu;
 	int ret = 0;
 
 	/*
@@ -5354,15 +5332,12 @@ again:
 	if (cpumask_test_cpu(task_cpu(p), new_mask))
 		goto out;
 
-	if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
+	dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
+	if (migrate_task(p, dest_cpu)) {
+		struct migration_arg arg = { p, dest_cpu };
 		/* Need help from migration thread: drop lock and wait. */
-		struct task_struct *mt = rq->migration_thread;
-
-		get_task_struct(mt);
 		task_rq_unlock(rq, &flags);
-		wake_up_process(mt);
-		put_task_struct(mt);
-		wait_for_completion(&req.done);
+		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
 		tlb_migrate_finish(p->mm);
 		return 0;
 	}
@@ -5420,70 +5395,22 @@ fail:
 	return ret;
 }
 
-#define RCU_MIGRATION_IDLE	0
-#define RCU_MIGRATION_NEED_QS	1
-#define RCU_MIGRATION_GOT_QS	2
-#define RCU_MIGRATION_MUST_SYNC	3
-
 /*
- * migration_thread - this is a highprio system thread that performs
- * thread migration by bumping thread off CPU then 'pushing' onto
- * another runqueue.
+ * migration_cpu_stop - this will be executed by a highprio stopper thread
+ * and performs thread migration by bumping thread off CPU then
+ * 'pushing' onto another runqueue.
  */
-static int migration_thread(void *data)
+static int migration_cpu_stop(void *data)
 {
-	int badcpu;
-	int cpu = (long)data;
-	struct rq *rq;
-
-	rq = cpu_rq(cpu);
-	BUG_ON(rq->migration_thread != current);
-
-	set_current_state(TASK_INTERRUPTIBLE);
-	while (!kthread_should_stop()) {
-		struct migration_req *req;
-		struct list_head *head;
-
-		raw_spin_lock_irq(&rq->lock);
-
-		if (cpu_is_offline(cpu)) {
-			raw_spin_unlock_irq(&rq->lock);
-			break;
-		}
-
-		if (rq->active_balance) {
-			active_load_balance(rq, cpu);
-			rq->active_balance = 0;
-		}
-
-		head = &rq->migration_queue;
-
-		if (list_empty(head)) {
-			raw_spin_unlock_irq(&rq->lock);
-			schedule();
-			set_current_state(TASK_INTERRUPTIBLE);
-			continue;
-		}
-		req = list_entry(head->next, struct migration_req, list);
-		list_del_init(head->next);
-
-		if (req->task != NULL) {
-			raw_spin_unlock(&rq->lock);
-			__migrate_task(req->task, cpu, req->dest_cpu);
-		} else if (likely(cpu == (badcpu = smp_processor_id()))) {
-			req->dest_cpu = RCU_MIGRATION_GOT_QS;
-			raw_spin_unlock(&rq->lock);
-		} else {
-			req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
-			raw_spin_unlock(&rq->lock);
-			WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
-		}
-		local_irq_enable();
-
-		complete(&req->done);
-	}
-	__set_current_state(TASK_RUNNING);
+	struct migration_arg *arg = data;
 
+	/*
+	 * The original target cpu might have gone down and we might
+	 * be on another cpu but it doesn't matter.
+	 */
+	local_irq_disable();
+	__migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
+	local_irq_enable();
 	return 0;
 }
 
@@ -5850,35 +5777,20 @@ static void set_rq_offline(struct rq *rq)
 static int __cpuinit
 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
 {
-	struct task_struct *p;
 	int cpu = (long)hcpu;
 	unsigned long flags;
-	struct rq *rq;
+	struct rq *rq = cpu_rq(cpu);
 
 	switch (action) {
 
 	case CPU_UP_PREPARE:
 	case CPU_UP_PREPARE_FROZEN:
-		p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
-		if (IS_ERR(p))
-			return NOTIFY_BAD;
-		kthread_bind(p, cpu);
-		/* Must be high prio: stop_machine expects to yield to it. */
-		rq = task_rq_lock(p, &flags);
-		__setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
-		task_rq_unlock(rq, &flags);
-		get_task_struct(p);
-		cpu_rq(cpu)->migration_thread = p;
 		rq->calc_load_update = calc_load_update;
 		break;
 
 	case CPU_ONLINE:
 	case CPU_ONLINE_FROZEN:
-		/* Strictly unnecessary, as first user will wake it. */
-		wake_up_process(cpu_rq(cpu)->migration_thread);
-
 		/* Update our root-domain */
-		rq = cpu_rq(cpu);
 		raw_spin_lock_irqsave(&rq->lock, flags);
 		if (rq->rd) {
 			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
@@ -5889,25 +5801,9 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
 		break;
 
 #ifdef CONFIG_HOTPLUG_CPU
-	case CPU_UP_CANCELED:
-	case CPU_UP_CANCELED_FROZEN:
-		if (!cpu_rq(cpu)->migration_thread)
-			break;
-		/* Unbind it from offline cpu so it can run. Fall thru. */
-		kthread_bind(cpu_rq(cpu)->migration_thread,
-			     cpumask_any(cpu_online_mask));
-		kthread_stop(cpu_rq(cpu)->migration_thread);
-		put_task_struct(cpu_rq(cpu)->migration_thread);
-		cpu_rq(cpu)->migration_thread = NULL;
-		break;
-
 	case CPU_DEAD:
 	case CPU_DEAD_FROZEN:
 		migrate_live_tasks(cpu);
-		rq = cpu_rq(cpu);
-		kthread_stop(rq->migration_thread);
-		put_task_struct(rq->migration_thread);
-		rq->migration_thread = NULL;
 		/* Idle task back to normal (off runqueue, low prio) */
 		raw_spin_lock_irq(&rq->lock);
 		deactivate_task(rq, rq->idle, 0);
@@ -5918,29 +5814,11 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
 		migrate_nr_uninterruptible(rq);
 		BUG_ON(rq->nr_running != 0);
 		calc_global_load_remove(rq);
-		/*
-		 * No need to migrate the tasks: it was best-effort if
-		 * they didn't take sched_hotcpu_mutex. Just wake up
-		 * the requestors.
-		 */
-		raw_spin_lock_irq(&rq->lock);
-		while (!list_empty(&rq->migration_queue)) {
-			struct migration_req *req;
-
-			req = list_entry(rq->migration_queue.next,
-					 struct migration_req, list);
-			list_del_init(&req->list);
-			raw_spin_unlock_irq(&rq->lock);
-			complete(&req->done);
-			raw_spin_lock_irq(&rq->lock);
-		}
-		raw_spin_unlock_irq(&rq->lock);
 		break;
 
 	case CPU_DYING:
 	case CPU_DYING_FROZEN:
 		/* Update our root-domain */
-		rq = cpu_rq(cpu);
 		raw_spin_lock_irqsave(&rq->lock, flags);
 		if (rq->rd) {
 			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
@@ -7757,10 +7635,8 @@ void __init sched_init(void)
 		rq->push_cpu = 0;
 		rq->cpu = i;
 		rq->online = 0;
-		rq->migration_thread = NULL;
 		rq->idle_stamp = 0;
 		rq->avg_idle = 2*sysctl_sched_migration_cost;
-		INIT_LIST_HEAD(&rq->migration_queue);
 		rq_attach_root(rq, &def_root_domain);
 #endif
 		init_rq_hrtick(rq);
@@ -9054,12 +8930,6 @@ struct cgroup_subsys cpuacct_subsys = {
 
 #ifndef CONFIG_SMP
 
-int rcu_expedited_torture_stats(char *page)
-{
-	return 0;
-}
-EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
-
 void synchronize_sched_expedited(void)
 {
 }
@@ -9067,30 +8937,20 @@ EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
 
 #else /* #ifndef CONFIG_SMP */
 
-static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
-static DEFINE_MUTEX(rcu_sched_expedited_mutex);
-
-#define RCU_EXPEDITED_STATE_POST -2
-#define RCU_EXPEDITED_STATE_IDLE -1
-
-static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
+static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
 
-int rcu_expedited_torture_stats(char *page)
+static int synchronize_sched_expedited_cpu_stop(void *data)
 {
-	int cnt = 0;
-	int cpu;
+	static DEFINE_SPINLOCK(done_mask_lock);
+	struct cpumask *done_mask = data;
 
-	cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
-	for_each_online_cpu(cpu) {
-		 cnt += sprintf(&page[cnt], " %d:%d",
-				cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
+	if (done_mask) {
+		spin_lock(&done_mask_lock);
+		cpumask_set_cpu(smp_processor_id(), done_mask);
+		spin_unlock(&done_mask_lock);
 	}
-	cnt += sprintf(&page[cnt], "\n");
-	return cnt;
+	return 0;
 }
-EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
-
-static long synchronize_sched_expedited_count;
 
 /*
  * Wait for an rcu-sched grace period to elapse, but use "big hammer"
@@ -9104,60 +8964,55 @@ static long synchronize_sched_expedited_count;
  */
 void synchronize_sched_expedited(void)
 {
-	int cpu;
-	unsigned long flags;
-	bool need_full_sync = 0;
-	struct rq *rq;
-	struct migration_req *req;
-	long snap;
-	int trycount = 0;
+	cpumask_var_t done_mask_var;
+	struct cpumask *done_mask = NULL;
+	int snap, trycount = 0;
+
+	/*
+	 * done_mask is used to check that all cpus actually have
+	 * finished running the stopper, which is guaranteed by
+	 * stop_cpus() if it's called with cpu hotplug blocked.  Keep
+	 * the paranoia for now but it's best effort if cpumask is off
+	 * stack.
+	 */
+	if (zalloc_cpumask_var(&done_mask_var, GFP_ATOMIC))
+		done_mask = done_mask_var;
 
 	smp_mb();  /* ensure prior mod happens before capturing snap. */
-	snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
+	snap = atomic_read(&synchronize_sched_expedited_count) + 1;
 	get_online_cpus();
-	while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
+	while (try_stop_cpus(cpu_online_mask,
+			     synchronize_sched_expedited_cpu_stop,
+			     done_mask) == -EAGAIN) {
 		put_online_cpus();
 		if (trycount++ < 10)
 			udelay(trycount * num_online_cpus());
 		else {
 			synchronize_sched();
-			return;
+			goto free_out;
 		}
-		if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
+		if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
 			smp_mb(); /* ensure test happens before caller kfree */
-			return;
+			goto free_out;
 		}
 		get_online_cpus();
 	}
-	rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
-	for_each_online_cpu(cpu) {
-		rq = cpu_rq(cpu);
-		req = &per_cpu(rcu_migration_req, cpu);
-		init_completion(&req->done);
-		req->task = NULL;
-		req->dest_cpu = RCU_MIGRATION_NEED_QS;
-		raw_spin_lock_irqsave(&rq->lock, flags);
-		list_add(&req->list, &rq->migration_queue);
-		raw_spin_unlock_irqrestore(&rq->lock, flags);
-		wake_up_process(rq->migration_thread);
-	}
-	for_each_online_cpu(cpu) {
-		rcu_expedited_state = cpu;
-		req = &per_cpu(rcu_migration_req, cpu);
-		rq = cpu_rq(cpu);
-		wait_for_completion(&req->done);
-		raw_spin_lock_irqsave(&rq->lock, flags);
-		if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
-			need_full_sync = 1;
-		req->dest_cpu = RCU_MIGRATION_IDLE;
-		raw_spin_unlock_irqrestore(&rq->lock, flags);
-	}
-	rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
-	synchronize_sched_expedited_count++;
-	mutex_unlock(&rcu_sched_expedited_mutex);
+	atomic_inc(&synchronize_sched_expedited_count);
+	if (done_mask)
+		cpumask_xor(done_mask, done_mask, cpu_online_mask);
 	put_online_cpus();
-	if (need_full_sync)
+
+	/* paranoia - this can't happen */
+	if (done_mask && cpumask_weight(done_mask)) {
+		char buf[80];
+
+		cpulist_scnprintf(buf, sizeof(buf), done_mask);
+		WARN_ONCE(1, "synchronize_sched_expedited: cpu online and done masks disagree on %d cpus: %s\n",
+			  cpumask_weight(done_mask), buf);
 		synchronize_sched();
+	}
+free_out:
+	free_cpumask_var(done_mask_var);
 }
 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
 
kernel/sched_fair.c +34 -14
index cbd8b8a..217e4a9 100644
@@ -2798,6 +2798,8 @@ static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle)
 	return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
 }
 
+static int active_load_balance_cpu_stop(void *data);
+
 /*
  * Check this_cpu to ensure it is balanced within domain. Attempt to move
  * tasks if there is an imbalance.
@@ -2887,8 +2889,9 @@ redo:
 		if (need_active_balance(sd, sd_idle, idle)) {
 			raw_spin_lock_irqsave(&busiest->lock, flags);
 
-			/* don't kick the migration_thread, if the curr
-			 * task on busiest cpu can't be moved to this_cpu
+			/* don't kick the active_load_balance_cpu_stop,
+			 * if the curr task on busiest cpu can't be
+			 * moved to this_cpu
 			 */
 			if (!cpumask_test_cpu(this_cpu,
 					      &busiest->curr->cpus_allowed)) {
@@ -2898,14 +2901,22 @@ redo:
 				goto out_one_pinned;
 			}
 
+			/*
+			 * ->active_balance synchronizes accesses to
+			 * ->active_balance_work.  Once set, it's cleared
+			 * only after active load balance is finished.
+			 */
 			if (!busiest->active_balance) {
 				busiest->active_balance = 1;
 				busiest->push_cpu = this_cpu;
 				active_balance = 1;
 			}
 			raw_spin_unlock_irqrestore(&busiest->lock, flags);
+
 			if (active_balance)
-				wake_up_process(busiest->migration_thread);
+				stop_one_cpu_nowait(cpu_of(busiest),
+					active_load_balance_cpu_stop, busiest,
+					&busiest->active_balance_work);
 
 			/*
 			 * We've kicked active balancing, reset the failure
@@ -3012,24 +3023,29 @@ static void idle_balance(int this_cpu, struct rq *this_rq)
 }
 
 /*
- * active_load_balance is run by migration threads. It pushes running tasks
- * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
- * running on each physical CPU where possible, and avoids physical /
- * logical imbalances.
- *
- * Called with busiest_rq locked.
+ * active_load_balance_cpu_stop is run by cpu stopper. It pushes
+ * running tasks off the busiest CPU onto idle CPUs. It requires at
+ * least 1 task to be running on each physical CPU where possible, and
+ * avoids physical / logical imbalances.
  */
-static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
+static int active_load_balance_cpu_stop(void *data)
 {
+	struct rq *busiest_rq = data;
+	int busiest_cpu = cpu_of(busiest_rq);
 	int target_cpu = busiest_rq->push_cpu;
+	struct rq *target_rq = cpu_rq(target_cpu);
 	struct sched_domain *sd;
-	struct rq *target_rq;
+
+	raw_spin_lock_irq(&busiest_rq->lock);
+
+	/* make sure the requested cpu hasn't gone down in the meantime */
+	if (unlikely(busiest_cpu != smp_processor_id() ||
+		     !busiest_rq->active_balance))
+		goto out_unlock;
 
 	/* Is there any task to move? */
 	if (busiest_rq->nr_running <= 1)
-		return;
-
-	target_rq = cpu_rq(target_cpu);
+		goto out_unlock;
 
 	/*
 	 * This condition is "impossible", if it occurs
@@ -3058,6 +3074,10 @@ static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
 			schedstat_inc(sd, alb_failed);
 	}
 	double_unlock_balance(busiest_rq, target_rq);
+out_unlock:
+	busiest_rq->active_balance = 0;
+	raw_spin_unlock_irq(&busiest_rq->lock);
+	return 0;
 }
 
 #ifdef CONFIG_NO_HZ
kernel/stop_machine.c +1 -1
index 884c7a1..5b20141 100644
@@ -301,7 +301,7 @@ static int __cpuinit cpu_stop_cpu_callback(struct notifier_block *nfb,
 	case CPU_UP_PREPARE:
 		BUG_ON(stopper->thread || stopper->enabled ||
 		       !list_empty(&stopper->works));
-		p = kthread_create(cpu_stopper_thread, stopper, "stopper/%d",
+		p = kthread_create(cpu_stopper_thread, stopper, "migration/%d",
 				   cpu);
 		if (IS_ERR(p))
 			return NOTIFY_BAD;
Signature · 1 line
1.6.4.2
author
#5 lore raw

Re: [PATCH 3/4] scheduler: replace migration_thread with cpu_stop

Hello,

Quoted 3 lines · Paul E. McKenney

o Therefore, when CPU 0 queues the work for CPU 1, CPU 1
loops right around and processes it. There will be no context switch on CPU 1.

Yes, that can happen.

Quoted 7 lines

At first glance, this looks safe because:

1. Although there is no context switch, there (presumably)
can be no RCU read-side critical sections on CPU 1 that span this sequence of events. (As far as I can see, any such RCU read-side critical section would be due to abuse of rcu_read_lock() and friends.)

AFAICS, this must hold; otherwise, synchronize_sched_expedited() wouldn't have worked in the first place. On entry to any cpu_stop function, there can be no RCU read-side critical section in progress.

Quoted 8 lines

2. CPU 1 will acquire and release stopper->lock, and
further more will do an atomic_dec_and_test() in cpu_stop_signal_done(). The former is a weak guarantee, but the latter guarantees a memory barrier, so that any subsequent code on CPU 1 will be guaranteed to see changes on CPU 0 prior to the call to synchronize_sched_expedited().

Quoted 4 lines

The guarantee required is that there will be a full memory barrier on each affected CPU between the time that try_stop_cpus() is called and the time that it returns.

Ah, right. I think it would be dangerous to depend on the implicit barriers there. It might work today but it can easily break with later implementation detail changes which seem completely unrelated. Adding smp_mb() in the cpu_stop function should suffice, right? It's not like the cost of smp_mb() there would mean anything anyway.

Thank you.

Signature · 1 line
tejun
#6 lore raw

Re: [PATCH 3/4] scheduler: replace migration_thread with cpu_stop

Quoted 17 lines · Tejun Heo

Hello,

Quoted 3 lines · Paul E. McKenney

o Therefore, when CPU 0 queues the work for CPU 1, CPU 1
loops right around and processes it. There will be no context switch on CPU 1.

Yes, that can happen.

Quoted 7 lines

At first glance, this looks safe because:

1. Although there is no context switch, there (presumably)
can be no RCU read-side critical sections on CPU 1 that span this sequence of events. (As far as I can see, any such RCU read-side critical section would be due to abuse of rcu_read_lock() and friends.)

Quoted 3 lines

AFAICS, this must hold; otherwise, synchronize_sched_expedited() wouldn't have worked in the first place. On entry to any cpu_stop function, there can be no RCU read-side critical section in progress.

Makes sense to me!

The actual requirement is that, on each CPU, there must have been a context switch between the end of the last RCU read-side critical section and the end of a successful return from try_stop_cpus().

For CONFIG_TREE_PREEMPT_RCU, the guarantee required is a bit different:
on each CPU, either that CPU must not have been in an RCU read-side critical section, or, if it was, there must have been a context switch between the time that CPU entered its RCU read-side critical section and the memory barrier executed within a successful try_stop_cpus().

As near as I can tell, the current implementation does meet these requirements (but I do like your suggested change below).

Quoted 13 lines
Quoted 8 lines

2. CPU 1 will acquire and release stopper->lock, and
further more will do an atomic_dec_and_test() in cpu_stop_signal_done(). The former is a weak guarantee, but the latter guarantees a memory barrier, so that any subsequent code on CPU 1 will be guaranteed to see changes on CPU 0 prior to the call to synchronize_sched_expedited().

Quoted 4 lines

The guarantee required is that there will be a full memory barrier on each affected CPU between the time that try_stop_cpus() is called and the time that it returns.

Quoted 5 lines

Ah, right. I think it would be dangerous to depend on the implicit barriers there. It might work today but it can easily break with later implementation detail changes which seem completely unrelated. Adding smp_mb() in the cpu_stop function should suffice, right? It's not like the cost of smp_mb() there would mean anything anyway.

If I understand the code correctly, this would be very good!

Thanx, Paul

author
#7 lore raw

PATCH 3/4 scheduler: replace migration_thread with cpu_stop

Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy.

This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary.

synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped.

synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree.

Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed.

This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes.

With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner.

Cc Ingo Molnar <[email protected]>
Cc Dipankar Sarma <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Paul E. McKenney <[email protected]>
Cc Oleg Nesterov <[email protected]>
Cc Dimitri Sivanich <[email protected]>

---
Alright, smp_mb() and comments added. Paul, can you please ack this?

Thank you.

 Documentation/RCU/torture.txt |   10 -
 include/linux/rcutiny.h       |    2
 include/linux/rcutree.h       |    1
 kernel/rcutorture.c           |    2
 kernel/sched.c                |  315 ++++++++++++------------------------------
 kernel/sched_fair.c           |   48 ++++--
 kernel/stop_machine.c         |    2
 7 files changed, 127 insertions(+), 253 deletions(-)
kernel/sched.c +91 -224
Index: work/kernel/sched.c
===================================================================
@@ -55,9 +55,9 @@
 #include <linux/cpu.h>
 #include <linux/cpuset.h>
 #include <linux/percpu.h>
-#include <linux/kthread.h>
 #include <linux/proc_fs.h>
 #include <linux/seq_file.h>
+#include <linux/stop_machine.h>
 #include <linux/sysctl.h>
 #include <linux/syscalls.h>
 #include <linux/times.h>
@@ -539,15 +539,13 @@ struct rq {
 	int post_schedule;
 	int active_balance;
 	int push_cpu;
+	struct cpu_stop_work active_balance_work;
 	/* cpu of this runqueue: */
 	int cpu;
 	int online;
 
 	unsigned long avg_load_per_task;
 
-	struct task_struct *migration_thread;
-	struct list_head migration_queue;
-
 	u64 rt_avg;
 	u64 age_stamp;
 	u64 idle_stamp;
@@ -2037,21 +2035,18 @@ void set_task_cpu(struct task_struct *p,
 	__set_task_cpu(p, new_cpu);
 }
 
-struct migration_req {
-	struct list_head list;
-
+struct migration_arg {
 	struct task_struct *task;
 	int dest_cpu;
-
-	struct completion done;
 };
 
+static int migration_cpu_stop(void *data);
+
 /*
  * The task's runqueue lock must be held.
  * Returns true if you have to wait for migration thread.
  */
-static int
-migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
+static bool migrate_task(struct task_struct *p, int dest_cpu)
 {
 	struct rq *rq = task_rq(p);
 
@@ -2059,15 +2054,7 @@ migrate_task(struct task_struct *p, int
 	 * If the task is not on a runqueue (and not running), then
 	 * the next wake-up will properly place the task.
 	 */
-	if (!p->se.on_rq && !task_running(rq, p))
-		return 0;
-
-	init_completion(&req->done);
-	req->task = p;
-	req->dest_cpu = dest_cpu;
-	list_add(&req->list, &rq->migration_queue);
-
-	return 1;
+	return p->se.on_rq || task_running(rq, p);
 }
 
 /*
@@ -3056,7 +3043,6 @@ static void update_cpu_load(struct rq *t
 void sched_exec(void)
 {
 	struct task_struct *p = current;
-	struct migration_req req;
 	unsigned long flags;
 	struct rq *rq;
 	int dest_cpu;
@@ -3070,17 +3056,11 @@ void sched_exec(void)
 	 * select_task_rq() can race against ->cpus_allowed
 	 */
 	if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
-	    likely(cpu_active(dest_cpu)) &&
-	    migrate_task(p, dest_cpu, &req)) {
-		/* Need to wait for migration thread (might exit: take ref). */
-		struct task_struct *mt = rq->migration_thread;
+	    likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
+		struct migration_arg arg = { p, dest_cpu };
 
-		get_task_struct(mt);
 		task_rq_unlock(rq, &flags);
-		wake_up_process(mt);
-		put_task_struct(mt);
-		wait_for_completion(&req.done);
-
+		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
 		return;
 	}
 unlock:
@@ -5236,17 +5216,15 @@ static inline void sched_init_granularit
 /*
  * This is how migration works:
  *
- * 1) we queue a struct migration_req structure in the source CPU's
- *    runqueue and wake up that CPU's migration thread.
- * 2) we down() the locked semaphore => thread blocks.
- * 3) migration thread wakes up (implicitly it forces the migrated
- *    thread off the CPU)
- * 4) it gets the migration request and checks whether the migrated
- *    task is still in the wrong runqueue.
- * 5) if it's in the wrong runqueue then the migration thread removes
+ * 1) we invoke migration_cpu_stop() on the target CPU using
+ *    stop_one_cpu().
+ * 2) stopper starts to run (implicitly forcing the migrated thread
+ *    off the CPU)
+ * 3) it checks whether the migrated task is still in the wrong runqueue.
+ * 4) if it's in the wrong runqueue then the migration thread removes
  *    it and puts it into the right queue.
- * 6) migration thread up()s the semaphore.
- * 7) we wake up and the migration is done.
+ * 5) stopper completes and stop_one_cpu() returns and the migration
+ *    is done.
  */
 
 /*
@@ -5260,9 +5238,9 @@ static inline void sched_init_granularit
  */
 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
 {
-	struct migration_req req;
 	unsigned long flags;
 	struct rq *rq;
+	unsigned int dest_cpu;
 	int ret = 0;
 
 	/*
@@ -5300,15 +5278,12 @@ again:
 	if (cpumask_test_cpu(task_cpu(p), new_mask))
 		goto out;
 
-	if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
+	dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
+	if (migrate_task(p, dest_cpu)) {
+		struct migration_arg arg = { p, dest_cpu };
 		/* Need help from migration thread: drop lock and wait. */
-		struct task_struct *mt = rq->migration_thread;
-
-		get_task_struct(mt);
 		task_rq_unlock(rq, &flags);
-		wake_up_process(mt);
-		put_task_struct(mt);
-		wait_for_completion(&req.done);
+		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
 		tlb_migrate_finish(p->mm);
 		return 0;
 	}
@@ -5366,70 +5341,22 @@ fail:
 	return ret;
 }
 
-#define RCU_MIGRATION_IDLE	0
-#define RCU_MIGRATION_NEED_QS	1
-#define RCU_MIGRATION_GOT_QS	2
-#define RCU_MIGRATION_MUST_SYNC	3
-
 /*
- * migration_thread - this is a highprio system thread that performs
- * thread migration by bumping thread off CPU then 'pushing' onto
- * another runqueue.
+ * migration_cpu_stop - this will be executed by a highprio stopper thread
+ * and performs thread migration by bumping thread off CPU then
+ * 'pushing' onto another runqueue.
  */
-static int migration_thread(void *data)
+static int migration_cpu_stop(void *data)
 {
-	int badcpu;
-	int cpu = (long)data;
-	struct rq *rq;
-
-	rq = cpu_rq(cpu);
-	BUG_ON(rq->migration_thread != current);
-
-	set_current_state(TASK_INTERRUPTIBLE);
-	while (!kthread_should_stop()) {
-		struct migration_req *req;
-		struct list_head *head;
-
-		raw_spin_lock_irq(&rq->lock);
-
-		if (cpu_is_offline(cpu)) {
-			raw_spin_unlock_irq(&rq->lock);
-			break;
-		}
-
-		if (rq->active_balance) {
-			active_load_balance(rq, cpu);
-			rq->active_balance = 0;
-		}
-
-		head = &rq->migration_queue;
-
-		if (list_empty(head)) {
-			raw_spin_unlock_irq(&rq->lock);
-			schedule();
-			set_current_state(TASK_INTERRUPTIBLE);
-			continue;
-		}
-		req = list_entry(head->next, struct migration_req, list);
-		list_del_init(head->next);
-
-		if (req->task != NULL) {
-			raw_spin_unlock(&rq->lock);
-			__migrate_task(req->task, cpu, req->dest_cpu);
-		} else if (likely(cpu == (badcpu = smp_processor_id()))) {
-			req->dest_cpu = RCU_MIGRATION_GOT_QS;
-			raw_spin_unlock(&rq->lock);
-		} else {
-			req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
-			raw_spin_unlock(&rq->lock);
-			WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
-		}
-		local_irq_enable();
-
-		complete(&req->done);
-	}
-	__set_current_state(TASK_RUNNING);
+	struct migration_arg *arg = data;
 
+	/*
+	 * The original target cpu might have gone down and we might
+	 * be on another cpu but it doesn't matter.
+	 */
+	local_irq_disable();
+	__migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
+	local_irq_enable();
 	return 0;
 }
 
@@ -5796,35 +5723,20 @@ static void set_rq_offline(struct rq *rq
 static int __cpuinit
 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
 {
-	struct task_struct *p;
 	int cpu = (long)hcpu;
 	unsigned long flags;
-	struct rq *rq;
+	struct rq *rq = cpu_rq(cpu);
 
 	switch (action) {
 
 	case CPU_UP_PREPARE:
 	case CPU_UP_PREPARE_FROZEN:
-		p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
-		if (IS_ERR(p))
-			return NOTIFY_BAD;
-		kthread_bind(p, cpu);
-		/* Must be high prio: stop_machine expects to yield to it. */
-		rq = task_rq_lock(p, &flags);
-		__setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
-		task_rq_unlock(rq, &flags);
-		get_task_struct(p);
-		cpu_rq(cpu)->migration_thread = p;
 		rq->calc_load_update = calc_load_update;
 		break;
 
 	case CPU_ONLINE:
 	case CPU_ONLINE_FROZEN:
-		/* Strictly unnecessary, as first user will wake it. */
-		wake_up_process(cpu_rq(cpu)->migration_thread);
-
 		/* Update our root-domain */
-		rq = cpu_rq(cpu);
 		raw_spin_lock_irqsave(&rq->lock, flags);
 		if (rq->rd) {
 			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
@@ -5835,25 +5747,9 @@ migration_call(struct notifier_block *nf
 		break;
 
 #ifdef CONFIG_HOTPLUG_CPU
-	case CPU_UP_CANCELED:
-	case CPU_UP_CANCELED_FROZEN:
-		if (!cpu_rq(cpu)->migration_thread)
-			break;
-		/* Unbind it from offline cpu so it can run. Fall thru. */
-		kthread_bind(cpu_rq(cpu)->migration_thread,
-			     cpumask_any(cpu_online_mask));
-		kthread_stop(cpu_rq(cpu)->migration_thread);
-		put_task_struct(cpu_rq(cpu)->migration_thread);
-		cpu_rq(cpu)->migration_thread = NULL;
-		break;
-
 	case CPU_DEAD:
 	case CPU_DEAD_FROZEN:
 		migrate_live_tasks(cpu);
-		rq = cpu_rq(cpu);
-		kthread_stop(rq->migration_thread);
-		put_task_struct(rq->migration_thread);
-		rq->migration_thread = NULL;
 		/* Idle task back to normal (off runqueue, low prio) */
 		raw_spin_lock_irq(&rq->lock);
 		deactivate_task(rq, rq->idle, 0);
@@ -5864,29 +5760,11 @@ migration_call(struct notifier_block *nf
 		migrate_nr_uninterruptible(rq);
 		BUG_ON(rq->nr_running != 0);
 		calc_global_load_remove(rq);
-		/*
-		 * No need to migrate the tasks: it was best-effort if
-		 * they didn't take sched_hotcpu_mutex. Just wake up
-		 * the requestors.
-		 */
-		raw_spin_lock_irq(&rq->lock);
-		while (!list_empty(&rq->migration_queue)) {
-			struct migration_req *req;
-
-			req = list_entry(rq->migration_queue.next,
-					 struct migration_req, list);
-			list_del_init(&req->list);
-			raw_spin_unlock_irq(&rq->lock);
-			complete(&req->done);
-			raw_spin_lock_irq(&rq->lock);
-		}
-		raw_spin_unlock_irq(&rq->lock);
 		break;
 
 	case CPU_DYING:
 	case CPU_DYING_FROZEN:
 		/* Update our root-domain */
-		rq = cpu_rq(cpu);
 		raw_spin_lock_irqsave(&rq->lock, flags);
 		if (rq->rd) {
 			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
@@ -7700,10 +7578,8 @@ void __init sched_init(void)
 		rq->push_cpu = 0;
 		rq->cpu = i;
 		rq->online = 0;
-		rq->migration_thread = NULL;
 		rq->idle_stamp = 0;
 		rq->avg_idle = 2*sysctl_sched_migration_cost;
-		INIT_LIST_HEAD(&rq->migration_queue);
 		rq_attach_root(rq, &def_root_domain);
 #endif
 		init_rq_hrtick(rq);
@@ -8997,43 +8873,39 @@ struct cgroup_subsys cpuacct_subsys = {
 
 #ifndef CONFIG_SMP
 
-int rcu_expedited_torture_stats(char *page)
-{
-	return 0;
-}
-EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
-
 void synchronize_sched_expedited(void)
 {
+	/*
+	 * There must be a full memory barrier on each affected CPU
+	 * between the time that try_stop_cpus() is called and the
+	 * time that it returns.
+	 *
+	 * In the current initial implementation of cpu_stop, the
+	 * above condition is already met when the control reaches
+	 * this point and the following smp_mb() is not strictly
+	 * necessary.  Do smp_mb() anyway for documentation and
+	 * robustness against future implementation changes.
+	 */
+	smp_mb();
 }
 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
 
 #else /* #ifndef CONFIG_SMP */
 
-static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
-static DEFINE_MUTEX(rcu_sched_expedited_mutex);
-
-#define RCU_EXPEDITED_STATE_POST -2
-#define RCU_EXPEDITED_STATE_IDLE -1
-
-static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
+static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
 
-int rcu_expedited_torture_stats(char *page)
+static int synchronize_sched_expedited_cpu_stop(void *data)
 {
-	int cnt = 0;
-	int cpu;
+	static DEFINE_SPINLOCK(done_mask_lock);
+	struct cpumask *done_mask = data;
 
-	cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
-	for_each_online_cpu(cpu) {
-		 cnt += sprintf(&page[cnt], " %d:%d",
-				cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
+	if (done_mask) {
+		spin_lock(&done_mask_lock);
+		cpumask_set_cpu(smp_processor_id(), done_mask);
+		spin_unlock(&done_mask_lock);
 	}
-	cnt += sprintf(&page[cnt], "\n");
-	return cnt;
+	return 0;
 }
-EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
-
-static long synchronize_sched_expedited_count;
 
 /*
  * Wait for an rcu-sched grace period to elapse, but use "big hammer"
@@ -9047,60 +8919,55 @@ static long synchronize_sched_expedited_
  */
 void synchronize_sched_expedited(void)
 {
-	int cpu;
-	unsigned long flags;
-	bool need_full_sync = 0;
-	struct rq *rq;
-	struct migration_req *req;
-	long snap;
-	int trycount = 0;
+	cpumask_var_t done_mask_var;
+	struct cpumask *done_mask = NULL;
+	int snap, trycount = 0;
+
+	/*
+	 * done_mask is used to check that all cpus actually have
+	 * finished running the stopper, which is guaranteed by
+	 * stop_cpus() if it's called with cpu hotplug blocked.  Keep
+	 * the paranoia for now but it's best effort if cpumask is off
+	 * stack.
+	 */
+	if (zalloc_cpumask_var(&done_mask_var, GFP_ATOMIC))
+		done_mask = done_mask_var;
 
 	smp_mb();  /* ensure prior mod happens before capturing snap. */
-	snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
+	snap = atomic_read(&synchronize_sched_expedited_count) + 1;
 	get_online_cpus();
-	while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
+	while (try_stop_cpus(cpu_online_mask,
+			     synchronize_sched_expedited_cpu_stop,
+			     done_mask) == -EAGAIN) {
 		put_online_cpus();
 		if (trycount++ < 10)
 			udelay(trycount * num_online_cpus());
 		else {
 			synchronize_sched();
-			return;
+			goto free_out;
 		}
-		if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
+		if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
 			smp_mb(); /* ensure test happens before caller kfree */
-			return;
+			goto free_out;
 		}
 		get_online_cpus();
 	}
-	rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
-	for_each_online_cpu(cpu) {
-		rq = cpu_rq(cpu);
-		req = &per_cpu(rcu_migration_req, cpu);
-		init_completion(&req->done);
-		req->task = NULL;
-		req->dest_cpu = RCU_MIGRATION_NEED_QS;
-		raw_spin_lock_irqsave(&rq->lock, flags);
-		list_add(&req->list, &rq->migration_queue);
-		raw_spin_unlock_irqrestore(&rq->lock, flags);
-		wake_up_process(rq->migration_thread);
-	}
-	for_each_online_cpu(cpu) {
-		rcu_expedited_state = cpu;
-		req = &per_cpu(rcu_migration_req, cpu);
-		rq = cpu_rq(cpu);
-		wait_for_completion(&req->done);
-		raw_spin_lock_irqsave(&rq->lock, flags);
-		if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
-			need_full_sync = 1;
-		req->dest_cpu = RCU_MIGRATION_IDLE;
-		raw_spin_unlock_irqrestore(&rq->lock, flags);
-	}
-	rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
-	synchronize_sched_expedited_count++;
-	mutex_unlock(&rcu_sched_expedited_mutex);
+	atomic_inc(&synchronize_sched_expedited_count);
+	if (done_mask)
+		cpumask_xor(done_mask, done_mask, cpu_online_mask);
 	put_online_cpus();
-	if (need_full_sync)
+
+	/* paranoia - this can't happen */
+	if (done_mask && cpumask_weight(done_mask)) {
+		char buf[80];
+
+		cpulist_scnprintf(buf, sizeof(buf), done_mask);
+		WARN_ONCE(1, "synchronize_sched_expedited: cpu online and done masks disagree on %d cpus: %s\n",
+			  cpumask_weight(done_mask), buf);
 		synchronize_sched();
+	}
+free_out:
+	free_cpumask_var(done_mask_var);
 }
 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
 
include/linux/rcutree.h +0 -1
Index: work/include/linux/rcutree.h
===================================================================
@@ -35,7 +35,6 @@ struct notifier_block;
 extern void rcu_sched_qs(int cpu);
 extern void rcu_bh_qs(int cpu);
 extern int rcu_needs_cpu(int cpu);
-extern int rcu_expedited_torture_stats(char *page);
 
 #ifdef CONFIG_TREE_PREEMPT_RCU
 
kernel/rcutorture.c +1 -1
Index: work/kernel/rcutorture.c
===================================================================
@@ -669,7 +669,7 @@ static struct rcu_torture_ops sched_expe
 	.sync		= synchronize_sched_expedited,
 	.cb_barrier	= NULL,
 	.fqs		= rcu_sched_force_quiescent_state,
-	.stats		= rcu_expedited_torture_stats,
+	.stats		= NULL,
 	.irq_capable	= 1,
 	.name		= "sched_expedited"
 };
Documentation/RCU/torture.txt +0 -10
Index: work/Documentation/RCU/torture.txt
===================================================================
@@ -182,16 +182,6 @@ Similarly, sched_expedited RCU provides
 	sched_expedited-torture: Reader Pipe:  12660320201 95875 0 0 0 0 0 0 0 0 0
 	sched_expedited-torture: Reader Batch:  12660424885 0 0 0 0 0 0 0 0 0 0
 	sched_expedited-torture: Free-Block Circulation:  1090795 1090795 1090794 1090793 1090792 1090791 1090790 1090789 1090788 1090787 0
-	state: -1 / 0:0 3:0 4:0
-
-As before, the first four lines are similar to those for RCU.
-The last line shows the task-migration state.  The first number is
--1 if synchronize_sched_expedited() is idle, -2 if in the process of
-posting wakeups to the migration kthreads, and N when waiting on CPU N.
-Each of the colon-separated fields following the "/" is a CPU:state pair.
-Valid states are "0" for idle, "1" for waiting for quiescent state,
-"2" for passed through quiescent state, and "3" when a race with a
-CPU-hotplug event forces use of the synchronize_sched() primitive.
 
 
 USAGE
kernel/sched_fair.c +34 -14

92 lines not on this page · raw message

include/linux/rcutiny.h +0 -2
Index: work/include/linux/rcutiny.h
===================================================================
@@ -60,8 +60,6 @@ static inline long rcu_batches_completed
 	return 0;
 }
 
-extern int rcu_expedited_torture_stats(char *page);
-
 static inline void rcu_force_quiescent_state(void)
 {
 }
kernel/stop_machine.c +1 -1
Index: work/kernel/stop_machine.c
===================================================================
@@ -301,7 +301,7 @@ static int __cpuinit cpu_stop_cpu_callba
 	case CPU_UP_PREPARE:
 		BUG_ON(stopper->thread || stopper->enabled ||
 		       !list_empty(&stopper->works));
-		p = kthread_create(cpu_stopper_thread, stopper, "stopper/%d",
+		p = kthread_create(cpu_stopper_thread, stopper, "migration/%d",
 				   cpu);
 		if (IS_ERR(p))
 			return NOTIFY_BAD;
#8 lore raw

Re: [PATCH 3/4 UPDATED] scheduler: replace migration_thread with cpu_stop

Quoted 47 lines · Tejun Heo

Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy.

This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary.

synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped.

synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree.

Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed.

This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes.

With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner.

Cc Ingo Molnar <[email protected]>
Cc Dipankar Sarma <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Paul E. McKenney <[email protected]>
Cc Oleg Nesterov <[email protected]>
Cc Dimitri Sivanich <[email protected]>
Quoted 2 lines

---
Alright, smp_mb() and comments added. Paul, can you please ack this?

Almost. With the following patch, I am good with it.

Thanx, Paul

------------------------------------------------------------------------

commit ab3e147752b1edab4588abd7a66f2505b7b433ed

Author: Paul E. McKenney <[email protected]>
Date:   Wed May 5 13:28:37 2010 -0700

sched: correctly place paranioa memory barriers in synchronize_sched_expedited()

The memory barriers must be in the SMP case, not in the !SMP case. Also add a barrier after the atomic_inc() in order to ensure that other CPUs see post-synchronize_sched_expedited() actions as following the expedited grace period.

kernel/sched.c +11 -10
index e9c6d79..155a16d 100644
@@ -8932,6 +8932,15 @@ struct cgroup_subsys cpuacct_subsys = {
 
 void synchronize_sched_expedited(void)
 {
+}
+EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
+
+#else /* #ifndef CONFIG_SMP */
+
+static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
+
+static int synchronize_sched_expedited_cpu_stop(void *data)
+{
 	/*
 	 * There must be a full memory barrier on each affected CPU
 	 * between the time that try_stop_cpus() is called and the
@@ -8943,16 +8952,7 @@ void synchronize_sched_expedited(void)
 	 * necessary.  Do smp_mb() anyway for documentation and
 	 * robustness against future implementation changes.
 	 */
-	smp_mb();
-}
-EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
-
-#else /* #ifndef CONFIG_SMP */
-
-static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
-
-static int synchronize_sched_expedited_cpu_stop(void *data)
-{
+	smp_mb(); /* See above comment block. */
 	return 0;
 }
 
@@ -8990,6 +8990,7 @@ void synchronize_sched_expedited(void)
 		get_online_cpus();
 	}
 	atomic_inc(&synchronize_sched_expedited_count);
+	smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
 	put_online_cpus();
 }
 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
author
#9 lore raw

Re: [PATCH 3/4 UPDATED] scheduler: replace migration_thread with cpu_stop

Quoted 1 line · Paul E. McKenney

Almost. With the following patch, I am good with it.

Ah, right. What was I thinking? I'll include your patch and push it to Ingo. Thanks.

Signature · 1 line
tejun
#10 lore raw

Re: [PATCH 3/4 UPDATED] scheduler: replace migration_thread with cpu_stop

Quoted 5 lines · Tejun Heo
Quoted 1 line · Paul E. McKenney

Almost. With the following patch, I am good with it.

Ah, right. What was I thinking? I'll include your patch and push it to Ingo. Thanks.

And of course, I overdid it in my patch when I removed smp_mb() from the !SMP version of synchronize_sched_expedited(). :-/

If synchronize_sched_expedited() is ever to be invoked from within kernel/sched.c in a UP kernel, there needs to be a "barrier()" in the !SMP version. So please see the following patch.

Thanx, Paul

commit 0eab52c0eeeb8507a83bdb37cc8f690b1c4f4a2c

Author: Paul E. McKenney <[email protected]>
Date:   Thu May 6 11:40:17 2010 -0700

rcu: need barrier() in UP synchronize_sched_expedited()

If synchronize_sched_expedited() is ever to be called from within kernel/sched.c in a !SMP PREEMPT kernel, the !SMP implementation needs a barrier().

kernel/sched.c +1 -0
index 155a16d..fbaf312 100644
@@ -8932,6 +8932,7 @@ struct cgroup_subsys cpuacct_subsys = {
 
 void synchronize_sched_expedited(void)
 {
+	barrier();
 }
 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
 
author
#11 lore raw

Re: [PATCH 3/4 UPDATED] scheduler: replace migration_thread with cpu_stop

Quoted 6 lines

And of course, I overdid it in my patch when I removed smp_mb() from the !SMP version of synchronize_sched_expedited(). :-/

If synchronize_sched_expedited() is ever to be invoked from within kernel/sched.c in a UP kernel, there needs to be a "barrier()" in the !SMP version. So please see the following patch.

Thanks. Pushed.

Signature · 1 line
tejun
author
#12 lore raw

PATCH 4/4 scheduler: kill paranoia check in synchronize_sched_expedited()

In reply to Tejun Heo

The paranoid check which verifies that the cpu_stop callback is actually called on all online cpus is completely superflous. It's guaranteed by cpu_stop facility and if it didn't work as advertised other things would go horribly wrong and trying to recover using synchronize_sched() wouldn't be very meaningful.

Kill the paranoid check. Removal of this feature is done as a separate step so that it can serve as a bisection point if something actually goes wrong.

Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Dipankar Sarma <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Paul E. McKenney <[email protected]>
Cc Oleg Nesterov <[email protected]>
Cc Dimitri Sivanich <[email protected]>
---
 kernel/sched.c |   40 +++-------------------------------------
 1 files changed, 3 insertions(+), 37 deletions(-)
kernel/sched.c +3 -37
index 8a198f1..c65ebbd 100644
@@ -8941,14 +8941,6 @@ static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
 
 static int synchronize_sched_expedited_cpu_stop(void *data)
 {
-	static DEFINE_SPINLOCK(done_mask_lock);
-	struct cpumask *done_mask = data;
-
-	if (done_mask) {
-		spin_lock(&done_mask_lock);
-		cpumask_set_cpu(smp_processor_id(), done_mask);
-		spin_unlock(&done_mask_lock);
-	}
 	return 0;
 }
 
@@ -8964,55 +8956,29 @@ static int synchronize_sched_expedited_cpu_stop(void *data)
  */
 void synchronize_sched_expedited(void)
 {
-	cpumask_var_t done_mask_var;
-	struct cpumask *done_mask = NULL;
 	int snap, trycount = 0;
 
-	/*
-	 * done_mask is used to check that all cpus actually have
-	 * finished running the stopper, which is guaranteed by
-	 * stop_cpus() if it's called with cpu hotplug blocked.  Keep
-	 * the paranoia for now but it's best effort if cpumask is off
-	 * stack.
-	 */
-	if (zalloc_cpumask_var(&done_mask_var, GFP_ATOMIC))
-		done_mask = done_mask_var;
-
 	smp_mb();  /* ensure prior mod happens before capturing snap. */
 	snap = atomic_read(&synchronize_sched_expedited_count) + 1;
 	get_online_cpus();
 	while (try_stop_cpus(cpu_online_mask,
 			     synchronize_sched_expedited_cpu_stop,
-			     done_mask) == -EAGAIN) {
+			     NULL) == -EAGAIN) {
 		put_online_cpus();
 		if (trycount++ < 10)
 			udelay(trycount * num_online_cpus());
 		else {
 			synchronize_sched();
-			goto free_out;
+			return;
 		}
 		if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
 			smp_mb(); /* ensure test happens before caller kfree */
-			goto free_out;
+			return;
 		}
 		get_online_cpus();
 	}
 	atomic_inc(&synchronize_sched_expedited_count);
-	if (done_mask)
-		cpumask_xor(done_mask, done_mask, cpu_online_mask);
 	put_online_cpus();
-
-	/* paranoia - this can't happen */
-	if (done_mask && cpumask_weight(done_mask)) {
-		char buf[80];
-
-		cpulist_scnprintf(buf, sizeof(buf), done_mask);
-		WARN_ONCE(1, "synchronize_sched_expedited: cpu online and done masks disagree on %d cpus: %s\n",
-			  cpumask_weight(done_mask), buf);
-		synchronize_sched();
-	}
-free_out:
-	free_cpumask_var(done_mask_var);
 }
 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
 
Signature · 1 line
1.6.4.2
author
#14 lore raw
Quoted 6 lines · Peter Zijlstra
Quoted 3 lines · Tejun Heo

Peter, if you ack the series, I'll add refresh the patches w/ your ACK and request pull into sched/core to Ingo.

Great, thank you. As soon as Paul's comment is resolved, I'll push it to Ingo.

Signature · 1 line
tejun