[patch v2 0/4] MTRR rendezvous deadlock fix and cleanups using stop_machine()

[patch v2 0/4] MTRR rendezvous deadlock fix and cleanups using stop_machine()

[Suresh Siddha]

Here is the v2 of the MTRR stop machine patchset.

First patch is a simple fix targeting 3.0 and the stable kernels.
This fixes the boot deadlock reported in the
https://bugzilla.novell.com/show_bug.cgi?id=672008

Second patch reorganizes the stop_cpus() code.

Third patch introduces stop_machine_from_inactive_cpu() so that we can do
stop machine from the cpu hotplug path, where the calling cpu is not
yet active.

Fourth patch uses the stop_machine() and stop_machine_from_inactive_cpu()
to implement the x86 MTRR rendezvous sequence and thus remove the duplicate
implementation of stop machine using stop_one_cpu_nowait().

changes from v1:
* update comments and use cpu active instead of the cpu online checks.

thanks,
suresh

[patch v2 1/4] x86, mtrr: lock stop machine during MTRR rendezvous sequence

[Suresh Siddha]

[-- Attachment #1: mtrr_stop_machine_quick_fix.patch --]
[-- Type: text/plain, Size: 4284 bytes --]

From: Suresh Siddha <suresh.b.siddha@intel.com>

MTRR rendezvous sequence using stop_one_cpu_nowait() can potentially
happen in parallel with another system wide rendezvous using
stop_machine(). This can lead to deadlock (The order in which
works are queued can be different on different cpu's. Some cpu's
will be running the first rendezvous handler and others will be running
the second rendezvous handler. Each set waiting for the other set to join
for the system wide rendezvous, leading to a deadlock).

MTRR rendezvous sequence is not implemented using stop_machine() as this
gets called both from the process context aswell as the cpu online paths
(where the cpu has not come online and the interrupts are disabled etc).
stop_machine() works with only online cpus.

For now, take the stop_machine mutex in the MTRR rendezvous sequence that
gets called from an online cpu (here we are in the process context
and can potentially sleep while taking the mutex). And the MTRR rendezvous
that gets triggered during cpu online doesn't need to take this stop_machine
lock (as the stop_machine() already ensures that there is no cpu hotplug
going on in parallel by doing get_online_cpus())

    TBD: Pursue a cleaner solution of extending the stop_machine()
         infrastructure to handle the case where the calling cpu is
         still not online and use this for MTRR rendezvous sequence.

fixes: https://bugzilla.novell.com/show_bug.cgi?id=672008

Reported-by: Vadim Kotelnikov <vadimuzzz@inbox.ru>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: stable@kernel.org # 2.6.35+, backport a week or two after this gets more testing in mainline
---
 arch/x86/kernel/cpu/mtrr/main.c |   23 +++++++++++++++++++++++
 include/linux/stop_machine.h    |    2 ++
 kernel/stop_machine.c           |    2 +-
 3 files changed, 26 insertions(+), 1 deletion(-)

Index: linux-2.6-tip/arch/x86/kernel/cpu/mtrr/main.c
===================================================================
--- linux-2.6-tip.orig/arch/x86/kernel/cpu/mtrr/main.c
+++ linux-2.6-tip/arch/x86/kernel/cpu/mtrr/main.c
@@ -248,6 +248,25 @@ set_mtrr(unsigned int reg, unsigned long
 	unsigned long flags;
 	int cpu;
 
+#ifdef CONFIG_SMP
+	/*
+	 * If this cpu is not yet active, we are in the cpu online path. There
+	 * can be no stop_machine() in parallel, as stop machine ensures this
+	 * by using get_online_cpus(). We can skip taking the stop_cpus_mutex,
+	 * as we don't need it and also we can't afford to block while waiting
+	 * for the mutex.
+	 *
+	 * If this cpu is active, we need to prevent stop_machine() happening
+	 * in parallel by taking the stop cpus mutex.
+	 *
+	 * Also, this is called in the context of cpu online path or in the
+	 * context where cpu hotplug is prevented. So checking the active status
+	 * of the raw_smp_processor_id() is safe.
+	 */
+	if (cpu_active(raw_smp_processor_id()))
+		mutex_lock(&stop_cpus_mutex);
+#endif
+
 	preempt_disable();
 
 	data.smp_reg = reg;
@@ -330,6 +349,10 @@ set_mtrr(unsigned int reg, unsigned long
 
 	local_irq_restore(flags);
 	preempt_enable();
+#ifdef CONFIG_SMP
+	if (cpu_active(raw_smp_processor_id()))
+		mutex_unlock(&stop_cpus_mutex);
+#endif
 }
 
 /**
Index: linux-2.6-tip/include/linux/stop_machine.h
===================================================================
--- linux-2.6-tip.orig/include/linux/stop_machine.h
+++ linux-2.6-tip/include/linux/stop_machine.h
@@ -27,6 +27,8 @@ struct cpu_stop_work {
 	struct cpu_stop_done	*done;
 };
 
+extern struct mutex stop_cpus_mutex;
+
 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);
Index: linux-2.6-tip/kernel/stop_machine.c
===================================================================
--- linux-2.6-tip.orig/kernel/stop_machine.c
+++ linux-2.6-tip/kernel/stop_machine.c
@@ -132,8 +132,8 @@ void stop_one_cpu_nowait(unsigned int cp
 	cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), work_buf);
 }
 
+DEFINE_MUTEX(stop_cpus_mutex);
 /* static data for stop_cpus */
-static DEFINE_MUTEX(stop_cpus_mutex);
 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
 
 int __stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)

[patch v2 2/4] stop_machine: reorganize stop_cpus() implementation

[Suresh Siddha]

[-- Attachment #1: reorganize_stop_cpus.patch --]
[-- Type: text/plain, Size: 2101 bytes --]

From: Tejun Heo <tj@kernel.org>

Refactor the queuing part of the stop cpus work from __stop_cpus() into
queue_stop_cpus_work().

The reorganization is to help future improvements to stop_machine()
and doesn't introduce any behavior difference.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 kernel/stop_machine.c |   16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)

Index: linux-2.6-tip/kernel/stop_machine.c
===================================================================
--- linux-2.6-tip.orig/kernel/stop_machine.c
+++ linux-2.6-tip/kernel/stop_machine.c
@@ -136,10 +136,11 @@ DEFINE_MUTEX(stop_cpus_mutex);
 /* static data for stop_cpus */
 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
 
-int __stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
+static void queue_stop_cpus_work(const struct cpumask *cpumask,
+				 cpu_stop_fn_t fn, void *arg,
+				 struct cpu_stop_done *done)
 {
 	struct cpu_stop_work *work;
-	struct cpu_stop_done done;
 	unsigned int cpu;
 
 	/* initialize works and done */
@@ -147,9 +148,8 @@ int __stop_cpus(const struct cpumask *cp
 		work = &per_cpu(stop_cpus_work, cpu);
 		work->fn = fn;
 		work->arg = arg;
-		work->done = &done;
+		work->done = done;
 	}
-	cpu_stop_init_done(&done, cpumask_weight(cpumask));
 
 	/*
 	 * Disable preemption while queueing to avoid getting
@@ -161,7 +161,15 @@ int __stop_cpus(const struct cpumask *cp
 		cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu),
 				    &per_cpu(stop_cpus_work, cpu));
 	preempt_enable();
+}
 
+static int __stop_cpus(const struct cpumask *cpumask,
+		       cpu_stop_fn_t fn, void *arg)
+{
+	struct cpu_stop_done done;
+
+	cpu_stop_init_done(&done, cpumask_weight(cpumask));
+	queue_stop_cpus_work(cpumask, fn, arg, &done);
 	wait_for_completion(&done.completion);
 	return done.executed ? done.ret : -ENOENT;
 }

[patch v2 3/4] stop_machine: implement stop_machine_from_inactive_cpu()

[Suresh Siddha]

[-- Attachment #1: implement_stop_machine_from_inactive_cpu.patch --]
[-- Type: text/plain, Size: 5276 bytes --]

From: Tejun Heo <tj@kernel.org>

Currently, mtrr wants stop_machine functionality while a CPU is being
brought up.  As stop_machine() requires the calling CPU to be active,
mtrr implements its own stop_machine using stop_one_cpu() on each
online CPU.  This doesn't only unnecessarily duplicate complex logic
but also introduces a possibility of deadlock when it races against
the generic stop_machine().

This patch implements stop_machine_from_inactive_cpu() to serve such
use cases.  Its functionality is basically the same as stop_machine();
however, it should be called from a CPU which isn't active and doesn't
depend on working scheduling on the calling CPU.

This is achieved by using busy loops for synchronization and
open-coding stop_cpus queuing and waiting with direct invocation of
fn() for local CPU inbetween.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 include/linux/stop_machine.h |   14 ++++++++-
 kernel/stop_machine.c        |   62 ++++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 73 insertions(+), 3 deletions(-)

Index: linux-2.6-tip/kernel/stop_machine.c
===================================================================
--- linux-2.6-tip.orig/kernel/stop_machine.c
+++ linux-2.6-tip/kernel/stop_machine.c
@@ -439,8 +439,15 @@ static int stop_machine_cpu_stop(void *d
 	struct stop_machine_data *smdata = data;
 	enum stopmachine_state curstate = STOPMACHINE_NONE;
 	int cpu = smp_processor_id(), err = 0;
+	unsigned long flags;
 	bool is_active;
 
+	/*
+	 * When called from stop_machine_from_inactive_cpu(), irq might
+	 * already be disabled.  Save the state and restore it on exit.
+	 */
+	local_save_flags(flags);
+
 	if (!smdata->active_cpus)
 		is_active = cpu == cpumask_first(cpu_online_mask);
 	else
@@ -468,7 +475,7 @@ static int stop_machine_cpu_stop(void *d
 		}
 	} while (curstate != STOPMACHINE_EXIT);
 
-	local_irq_enable();
+	local_irq_restore(flags);
 	return err;
 }
 
@@ -495,4 +502,57 @@ int stop_machine(int (*fn)(void *), void
 }
 EXPORT_SYMBOL_GPL(stop_machine);
 
+/**
+ * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
+ * @fn: the function to run
+ * @data: the data ptr for the @fn()
+ * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
+ *
+ * This is identical to stop_machine() but can be called from a CPU which
+ * is not active.  The local CPU is in the process of hotplug (so no other
+ * CPU hotplug can start) and not marked active and doesn't have enough
+ * context to sleep.
+ *
+ * This function provides stop_machine() functionality for such state by
+ * using busy-wait for synchronization and executing @fn directly for local
+ * CPU.
+ *
+ * CONTEXT:
+ * Local CPU is inactive.  Temporarily stops all active CPUs.
+ *
+ * RETURNS:
+ * 0 if all executions of @fn returned 0, any non zero return value if any
+ * returned non zero.
+ */
+int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
+				  const struct cpumask *cpus)
+{
+	struct stop_machine_data smdata = { .fn = fn, .data = data,
+					    .active_cpus = cpus };
+	struct cpu_stop_done done;
+	int ret;
+
+	/* Local CPU must be inactive and CPU hotplug in progress. */
+	BUG_ON(cpu_active(raw_smp_processor_id()));
+	smdata.num_threads = num_active_cpus() + 1;	/* +1 for local */
+
+	/* No proper task established and can't sleep - busy wait for lock. */
+	while (!mutex_trylock(&stop_cpus_mutex))
+		cpu_relax();
+
+	/* Schedule work on other CPUs and execute directly for local CPU */
+	set_state(&smdata, STOPMACHINE_PREPARE);
+	cpu_stop_init_done(&done, num_active_cpus());
+	queue_stop_cpus_work(cpu_active_mask, stop_machine_cpu_stop, &smdata,
+			     &done);
+	ret = stop_machine_cpu_stop(&smdata);
+
+	/* Busy wait for completion. */
+	while (!completion_done(&done.completion))
+		cpu_relax();
+
+	mutex_unlock(&stop_cpus_mutex);
+	return ret ?: done.ret;
+}
+
 #endif	/* CONFIG_STOP_MACHINE */
Index: linux-2.6-tip/include/linux/stop_machine.h
===================================================================
--- linux-2.6-tip.orig/include/linux/stop_machine.h
+++ linux-2.6-tip/include/linux/stop_machine.h
@@ -126,15 +126,19 @@ int stop_machine(int (*fn)(void *), void
  */
 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus);
 
+int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
+				   const struct cpumask *cpus);
+
 #else	 /* CONFIG_STOP_MACHINE && CONFIG_SMP */
 
 static inline int __stop_machine(int (*fn)(void *), void *data,
 				 const struct cpumask *cpus)
 {
+	unsigned long flags;
 	int ret;
-	local_irq_disable();
+	local_irq_save(flags);
 	ret = fn(data);
-	local_irq_enable();
+	local_irq_restore(flags);
 	return ret;
 }
 
@@ -144,5 +148,11 @@ static inline int stop_machine(int (*fn)
 	return __stop_machine(fn, data, cpus);
 }
 
+static inline int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
+						 const struct cpumask *cpus)
+{
+	return __stop_machine(fn, data, cpus);
+}
+
 #endif	/* CONFIG_STOP_MACHINE && CONFIG_SMP */
 #endif	/* _LINUX_STOP_MACHINE */

[patch v2 4/4] x86, mtrr: use stop_machine APIs for doing MTRR rendezvous

[Suresh Siddha]

[-- Attachment #1: use_stop_machine_for_mtrr_rendezvous.patch --]
[-- Type: text/plain, Size: 9947 bytes --]

From: Suresh Siddha <suresh.b.siddha@intel.com>

MTRR rendezvous sequence is not implemened using stop_machine() before, as this
gets called both from the process context aswell as the cpu online paths
(where the cpu has not come online and the interrupts are disabled etc).

Now that we have a new stop_machine_from_inactive_cpu() API, use it for
rendezvous during mtrr init of a logical processor that is coming online.

For the rest (runtime MTRR modification, system boot, resume paths), use
stop_machine() to implement the rendezvous sequence. This will consolidate and
cleanup the code.

Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
---
 arch/x86/kernel/cpu/mtrr/main.c |  194 ++++++++--------------------------------
 include/linux/stop_machine.h    |    2 
 kernel/stop_machine.c           |    2 
 3 files changed, 43 insertions(+), 155 deletions(-)

Index: linux-2.6-tip/arch/x86/kernel/cpu/mtrr/main.c
===================================================================
--- linux-2.6-tip.orig/arch/x86/kernel/cpu/mtrr/main.c
+++ linux-2.6-tip/arch/x86/kernel/cpu/mtrr/main.c
@@ -137,55 +137,43 @@ static void __init init_table(void)
 }
 
 struct set_mtrr_data {
-	atomic_t	count;
-	atomic_t	gate;
 	unsigned long	smp_base;
 	unsigned long	smp_size;
 	unsigned int	smp_reg;
 	mtrr_type	smp_type;
 };
 
-static DEFINE_PER_CPU(struct cpu_stop_work, mtrr_work);
-
 /**
- * mtrr_work_handler - Synchronisation handler. Executed by "other" CPUs.
+ * mtrr_rendezvous_handler - Work done in the synchronization handler. Executed
+ * by all the CPUs.
  * @info: pointer to mtrr configuration data
  *
  * Returns nothing.
  */
-static int mtrr_work_handler(void *info)
+static int mtrr_rendezvous_handler(void *info)
 {
 #ifdef CONFIG_SMP
 	struct set_mtrr_data *data = info;
-	unsigned long flags;
-
-	atomic_dec(&data->count);
-	while (!atomic_read(&data->gate))
-		cpu_relax();
-
-	local_irq_save(flags);
-
-	atomic_dec(&data->count);
-	while (atomic_read(&data->gate))
-		cpu_relax();
 
-	/*  The master has cleared me to execute  */
+	/*
+	 * We use this same function to initialize the mtrrs during boot,
+	 * resume, runtime cpu online and on an explicit request to set a
+	 * specific MTRR.
+	 *
+	 * During boot or suspend, the state of the boot cpu's mtrrs has been
+	 * saved, and we want to replicate that across all the cpus that come
+	 * online (either at the end of boot or resume or during a runtime cpu
+	 * online). If we're doing that, @reg is set to something special and on
+	 * all the cpu's we do mtrr_if->set_all() (On the logical cpu that
+	 * started the boot/resume sequence, this might be a duplicate
+	 * set_all()).
+	 */
 	if (data->smp_reg != ~0U) {
 		mtrr_if->set(data->smp_reg, data->smp_base,
 			     data->smp_size, data->smp_type);
-	} else if (mtrr_aps_delayed_init) {
-		/*
-		 * Initialize the MTRRs inaddition to the synchronisation.
-		 */
+	} else if (mtrr_aps_delayed_init || !cpu_online(smp_processor_id())) {
 		mtrr_if->set_all();
 	}
-
-	atomic_dec(&data->count);
-	while (!atomic_read(&data->gate))
-		cpu_relax();
-
-	atomic_dec(&data->count);
-	local_irq_restore(flags);
 #endif
 	return 0;
 }
@@ -223,20 +211,11 @@ static inline int types_compatible(mtrr_
  * 14. Wait for buddies to catch up
  * 15. Enable interrupts.
  *
- * What does that mean for us? Well, first we set data.count to the number
- * of CPUs. As each CPU announces that it started the rendezvous handler by
- * decrementing the count, We reset data.count and set the data.gate flag
- * allowing all the cpu's to proceed with the work. As each cpu disables
- * interrupts, it'll decrement data.count once. We wait until it hits 0 and
- * proceed. We clear the data.gate flag and reset data.count. Meanwhile, they
- * are waiting for that flag to be cleared. Once it's cleared, each
- * CPU goes through the transition of updating MTRRs.
- * The CPU vendors may each do it differently,
- * so we call mtrr_if->set() callback and let them take care of it.
- * When they're done, they again decrement data->count and wait for data.gate
- * to be set.
- * When we finish, we wait for data.count to hit 0 and toggle the data.gate flag
- * Everyone then enables interrupts and we all continue on.
+ * What does that mean for us? Well, stop_machine() will ensure that
+ * the rendezvous handler is started on each CPU. And in lockstep they
+ * do the state transition of disabling interrupts, updating MTRR's
+ * (the CPU vendors may each do it differently, so we call mtrr_if->set()
+ * callback and let them take care of it.) and enabling interrupts.
  *
  * Note that the mechanism is the same for UP systems, too; all the SMP stuff
  * becomes nops.
@@ -244,115 +223,26 @@ static inline int types_compatible(mtrr_
 static void
 set_mtrr(unsigned int reg, unsigned long base, unsigned long size, mtrr_type type)
 {
-	struct set_mtrr_data data;
-	unsigned long flags;
-	int cpu;
-
-#ifdef CONFIG_SMP
-	/*
-	 * If this cpu is not yet active, we are in the cpu online path. There
-	 * can be no stop_machine() in parallel, as stop machine ensures this
-	 * by using get_online_cpus(). We can skip taking the stop_cpus_mutex,
-	 * as we don't need it and also we can't afford to block while waiting
-	 * for the mutex.
-	 *
-	 * If this cpu is active, we need to prevent stop_machine() happening
-	 * in parallel by taking the stop cpus mutex.
-	 *
-	 * Also, this is called in the context of cpu online path or in the
-	 * context where cpu hotplug is prevented. So checking the active status
-	 * of the raw_smp_processor_id() is safe.
-	 */
-	if (cpu_active(raw_smp_processor_id()))
-		mutex_lock(&stop_cpus_mutex);
-#endif
-
-	preempt_disable();
-
-	data.smp_reg = reg;
-	data.smp_base = base;
-	data.smp_size = size;
-	data.smp_type = type;
-	atomic_set(&data.count, num_booting_cpus() - 1);
-
-	/* Make sure data.count is visible before unleashing other CPUs */
-	smp_wmb();
-	atomic_set(&data.gate, 0);
-
-	/* Start the ball rolling on other CPUs */
-	for_each_online_cpu(cpu) {
-		struct cpu_stop_work *work = &per_cpu(mtrr_work, cpu);
-
-		if (cpu == smp_processor_id())
-			continue;
+	struct set_mtrr_data data = { .smp_reg = reg,
+				      .smp_base = base,
+				      .smp_size = size,
+				      .smp_type = type
+				    };
+
+	stop_machine(mtrr_rendezvous_handler, &data, cpu_online_mask);
+}
+
+static void set_mtrr_from_inactive_cpu(unsigned int reg, unsigned long base,
+				      unsigned long size, mtrr_type type)
+{
+	struct set_mtrr_data data = { .smp_reg = reg,
+				      .smp_base = base,
+				      .smp_size = size,
+				      .smp_type = type
+				    };
 
-		stop_one_cpu_nowait(cpu, mtrr_work_handler, &data, work);
-	}
-
-
-	while (atomic_read(&data.count))
-		cpu_relax();
-
-	/* Ok, reset count and toggle gate */
-	atomic_set(&data.count, num_booting_cpus() - 1);
-	smp_wmb();
-	atomic_set(&data.gate, 1);
-
-	local_irq_save(flags);
-
-	while (atomic_read(&data.count))
-		cpu_relax();
-
-	/* Ok, reset count and toggle gate */
-	atomic_set(&data.count, num_booting_cpus() - 1);
-	smp_wmb();
-	atomic_set(&data.gate, 0);
-
-	/* Do our MTRR business */
-
-	/*
-	 * HACK!
-	 *
-	 * We use this same function to initialize the mtrrs during boot,
-	 * resume, runtime cpu online and on an explicit request to set a
-	 * specific MTRR.
-	 *
-	 * During boot or suspend, the state of the boot cpu's mtrrs has been
-	 * saved, and we want to replicate that across all the cpus that come
-	 * online (either at the end of boot or resume or during a runtime cpu
-	 * online). If we're doing that, @reg is set to something special and on
-	 * this cpu we still do mtrr_if->set_all(). During boot/resume, this
-	 * is unnecessary if at this point we are still on the cpu that started
-	 * the boot/resume sequence. But there is no guarantee that we are still
-	 * on the same cpu. So we do mtrr_if->set_all() on this cpu aswell to be
-	 * sure that we are in sync with everyone else.
-	 */
-	if (reg != ~0U)
-		mtrr_if->set(reg, base, size, type);
-	else
-		mtrr_if->set_all();
-
-	/* Wait for the others */
-	while (atomic_read(&data.count))
-		cpu_relax();
-
-	atomic_set(&data.count, num_booting_cpus() - 1);
-	smp_wmb();
-	atomic_set(&data.gate, 1);
-
-	/*
-	 * Wait here for everyone to have seen the gate change
-	 * So we're the last ones to touch 'data'
-	 */
-	while (atomic_read(&data.count))
-		cpu_relax();
-
-	local_irq_restore(flags);
-	preempt_enable();
-#ifdef CONFIG_SMP
-	if (cpu_active(raw_smp_processor_id()))
-		mutex_unlock(&stop_cpus_mutex);
-#endif
+	stop_machine_from_inactive_cpu(mtrr_rendezvous_handler, &data,
+				       cpu_callout_mask);
 }
 
 /**
@@ -806,7 +696,7 @@ void mtrr_ap_init(void)
 	 *   2. cpu hotadd time. We let mtrr_add/del_page hold cpuhotplug
 	 *      lock to prevent mtrr entry changes
 	 */
-	set_mtrr(~0U, 0, 0, 0);
+	set_mtrr_from_inactive_cpu(~0U, 0, 0, 0);
 }
 
 /**
Index: linux-2.6-tip/include/linux/stop_machine.h
===================================================================
--- linux-2.6-tip.orig/include/linux/stop_machine.h
+++ linux-2.6-tip/include/linux/stop_machine.h
@@ -27,8 +27,6 @@ struct cpu_stop_work {
 	struct cpu_stop_done	*done;
 };
 
-extern struct mutex stop_cpus_mutex;
-
 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);
Index: linux-2.6-tip/kernel/stop_machine.c
===================================================================
--- linux-2.6-tip.orig/kernel/stop_machine.c
+++ linux-2.6-tip/kernel/stop_machine.c
@@ -132,8 +132,8 @@ void stop_one_cpu_nowait(unsigned int cp
 	cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), work_buf);
 }
 
-DEFINE_MUTEX(stop_cpus_mutex);
 /* static data for stop_cpus */
+static DEFINE_MUTEX(stop_cpus_mutex);
 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
 
 static void queue_stop_cpus_work(const struct cpumask *cpumask,

[tip:x86/urgent] x86, mtrr: lock stop machine during MTRR rendezvous sequence

[tip-bot for Suresh Siddha]

Commit-ID:  6d3321e8e2b3bf6a5892e2ef673c7bf536e3f904
Gitweb:     http://git.kernel.org/tip/6d3321e8e2b3bf6a5892e2ef673c7bf536e3f904
Author:     Suresh Siddha <suresh.b.siddha@intel.com>
AuthorDate: Thu, 23 Jun 2011 11:19:26 -0700
Committer:  H. Peter Anvin <hpa@linux.intel.com>
CommitDate: Mon, 27 Jun 2011 14:00:46 -0700

x86, mtrr: lock stop machine during MTRR rendezvous sequence

MTRR rendezvous sequence using stop_one_cpu_nowait() can potentially
happen in parallel with another system wide rendezvous using
stop_machine(). This can lead to deadlock (The order in which
works are queued can be different on different cpu's. Some cpu's
will be running the first rendezvous handler and others will be running
the second rendezvous handler. Each set waiting for the other set to join
for the system wide rendezvous, leading to a deadlock).

MTRR rendezvous sequence is not implemented using stop_machine() as this
gets called both from the process context aswell as the cpu online paths
(where the cpu has not come online and the interrupts are disabled etc).
stop_machine() works with only online cpus.

For now, take the stop_machine mutex in the MTRR rendezvous sequence that
gets called from an online cpu (here we are in the process context
and can potentially sleep while taking the mutex). And the MTRR rendezvous
that gets triggered during cpu online doesn't need to take this stop_machine
lock (as the stop_machine() already ensures that there is no cpu hotplug
going on in parallel by doing get_online_cpus())

    TBD: Pursue a cleaner solution of extending the stop_machine()
         infrastructure to handle the case where the calling cpu is
         still not online and use this for MTRR rendezvous sequence.

fixes: https://bugzilla.novell.com/show_bug.cgi?id=672008

Reported-by: Vadim Kotelnikov <vadimuzzz@inbox.ru>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Link: http://lkml.kernel.org/r/20110623182056.807230326@sbsiddha-MOBL3.sc.intel.com
Cc: stable@kernel.org # 2.6.35+, backport a week or two after this gets more testing in mainline
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
---
 arch/x86/kernel/cpu/mtrr/main.c |   23 +++++++++++++++++++++++
 include/linux/stop_machine.h    |    2 ++
 kernel/stop_machine.c           |    2 +-
 3 files changed, 26 insertions(+), 1 deletions(-)

diff --git a/arch/x86/kernel/cpu/mtrr/main.c b/arch/x86/kernel/cpu/mtrr/main.c
index 929739a..3d17bc7 100644
--- a/arch/x86/kernel/cpu/mtrr/main.c
+++ b/arch/x86/kernel/cpu/mtrr/main.c
@@ -248,6 +248,25 @@ set_mtrr(unsigned int reg, unsigned long base, unsigned long size, mtrr_type typ
 	unsigned long flags;
 	int cpu;
 
+#ifdef CONFIG_SMP
+	/*
+	 * If this cpu is not yet active, we are in the cpu online path. There
+	 * can be no stop_machine() in parallel, as stop machine ensures this
+	 * by using get_online_cpus(). We can skip taking the stop_cpus_mutex,
+	 * as we don't need it and also we can't afford to block while waiting
+	 * for the mutex.
+	 *
+	 * If this cpu is active, we need to prevent stop_machine() happening
+	 * in parallel by taking the stop cpus mutex.
+	 *
+	 * Also, this is called in the context of cpu online path or in the
+	 * context where cpu hotplug is prevented. So checking the active status
+	 * of the raw_smp_processor_id() is safe.
+	 */
+	if (cpu_active(raw_smp_processor_id()))
+		mutex_lock(&stop_cpus_mutex);
+#endif
+
 	preempt_disable();
 
 	data.smp_reg = reg;
@@ -330,6 +349,10 @@ set_mtrr(unsigned int reg, unsigned long base, unsigned long size, mtrr_type typ
 
 	local_irq_restore(flags);
 	preempt_enable();
+#ifdef CONFIG_SMP
+	if (cpu_active(raw_smp_processor_id()))
+		mutex_unlock(&stop_cpus_mutex);
+#endif
 }
 
 /**
diff --git a/include/linux/stop_machine.h b/include/linux/stop_machine.h
index 092dc9b..14d3524 100644
--- a/include/linux/stop_machine.h
+++ b/include/linux/stop_machine.h
@@ -27,6 +27,8 @@ struct cpu_stop_work {
 	struct cpu_stop_done	*done;
 };
 
+extern struct mutex stop_cpus_mutex;
+
 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);
diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c
index e3516b2..0cae1cc 100644
--- a/kernel/stop_machine.c
+++ b/kernel/stop_machine.c
@@ -132,8 +132,8 @@ void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
 	cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), work_buf);
 }
 
+DEFINE_MUTEX(stop_cpus_mutex);
 /* static data for stop_cpus */
-static DEFINE_MUTEX(stop_cpus_mutex);
 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
 
 int __stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)

[tip:x86/mtrr] stop_machine: reorganize stop_cpus() implementation

[tip-bot for Tejun Heo]

Commit-ID:  fd7355ba1e936487f5aae6fc058c6cb300e44a64
Gitweb:     http://git.kernel.org/tip/fd7355ba1e936487f5aae6fc058c6cb300e44a64
Author:     Tejun Heo <tj@kernel.org>
AuthorDate: Thu, 23 Jun 2011 11:19:27 -0700
Committer:  H. Peter Anvin <hpa@linux.intel.com>
CommitDate: Mon, 27 Jun 2011 15:17:07 -0700

stop_machine: reorganize stop_cpus() implementation

Refactor the queuing part of the stop cpus work from __stop_cpus() into
queue_stop_cpus_work().

The reorganization is to help future improvements to stop_machine()
and doesn't introduce any behavior difference.

Signed-off-by: Tejun Heo <tj@kernel.org>
Link: http://lkml.kernel.org/r/20110623182056.897818337@sbsiddha-MOBL3.sc.intel.com
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
---
 kernel/stop_machine.c |   16 ++++++++++++----
 1 files changed, 12 insertions(+), 4 deletions(-)

diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c
index 0cae1cc..4c89ee9 100644
--- a/kernel/stop_machine.c
+++ b/kernel/stop_machine.c
@@ -136,10 +136,11 @@ DEFINE_MUTEX(stop_cpus_mutex);
 /* static data for stop_cpus */
 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
 
-int __stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
+static void queue_stop_cpus_work(const struct cpumask *cpumask,
+				 cpu_stop_fn_t fn, void *arg,
+				 struct cpu_stop_done *done)
 {
 	struct cpu_stop_work *work;
-	struct cpu_stop_done done;
 	unsigned int cpu;
 
 	/* initialize works and done */
@@ -147,9 +148,8 @@ int __stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
 		work = &per_cpu(stop_cpus_work, cpu);
 		work->fn = fn;
 		work->arg = arg;
-		work->done = &done;
+		work->done = done;
 	}
-	cpu_stop_init_done(&done, cpumask_weight(cpumask));
 
 	/*
 	 * Disable preemption while queueing to avoid getting
@@ -161,7 +161,15 @@ int __stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
 		cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu),
 				    &per_cpu(stop_cpus_work, cpu));
 	preempt_enable();
+}
 
+static int __stop_cpus(const struct cpumask *cpumask,
+		       cpu_stop_fn_t fn, void *arg)
+{
+	struct cpu_stop_done done;
+
+	cpu_stop_init_done(&done, cpumask_weight(cpumask));
+	queue_stop_cpus_work(cpumask, fn, arg, &done);
 	wait_for_completion(&done.completion);
 	return done.executed ? done.ret : -ENOENT;
 }

[tip:x86/mtrr] stop_machine: implement stop_machine_from_inactive_cpu()

[tip-bot for Tejun Heo]

Commit-ID:  f740e6cd0cb5e7468e46831aeb4d9c30e03d5ebc
Gitweb:     http://git.kernel.org/tip/f740e6cd0cb5e7468e46831aeb4d9c30e03d5ebc
Author:     Tejun Heo <tj@kernel.org>
AuthorDate: Thu, 23 Jun 2011 11:19:28 -0700
Committer:  H. Peter Anvin <hpa@linux.intel.com>
CommitDate: Mon, 27 Jun 2011 15:17:08 -0700

stop_machine: implement stop_machine_from_inactive_cpu()

Currently, mtrr wants stop_machine functionality while a CPU is being
brought up.  As stop_machine() requires the calling CPU to be active,
mtrr implements its own stop_machine using stop_one_cpu() on each
online CPU.  This doesn't only unnecessarily duplicate complex logic
but also introduces a possibility of deadlock when it races against
the generic stop_machine().

This patch implements stop_machine_from_inactive_cpu() to serve such
use cases.  Its functionality is basically the same as stop_machine();
however, it should be called from a CPU which isn't active and doesn't
depend on working scheduling on the calling CPU.

This is achieved by using busy loops for synchronization and
open-coding stop_cpus queuing and waiting with direct invocation of
fn() for local CPU inbetween.

Signed-off-by: Tejun Heo <tj@kernel.org>
Link: http://lkml.kernel.org/r/20110623182056.982526827@sbsiddha-MOBL3.sc.intel.com
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
---
 include/linux/stop_machine.h |   14 ++++++++-
 kernel/stop_machine.c        |   62 +++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 73 insertions(+), 3 deletions(-)

diff --git a/include/linux/stop_machine.h b/include/linux/stop_machine.h
index 14d3524..e0f2da2 100644
--- a/include/linux/stop_machine.h
+++ b/include/linux/stop_machine.h
@@ -126,15 +126,19 @@ int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus);
  */
 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus);
 
+int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
+				   const struct cpumask *cpus);
+
 #else	 /* CONFIG_STOP_MACHINE && CONFIG_SMP */
 
 static inline int __stop_machine(int (*fn)(void *), void *data,
 				 const struct cpumask *cpus)
 {
+	unsigned long flags;
 	int ret;
-	local_irq_disable();
+	local_irq_save(flags);
 	ret = fn(data);
-	local_irq_enable();
+	local_irq_restore(flags);
 	return ret;
 }
 
@@ -144,5 +148,11 @@ static inline int stop_machine(int (*fn)(void *), void *data,
 	return __stop_machine(fn, data, cpus);
 }
 
+static inline int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
+						 const struct cpumask *cpus)
+{
+	return __stop_machine(fn, data, cpus);
+}
+
 #endif	/* CONFIG_STOP_MACHINE && CONFIG_SMP */
 #endif	/* _LINUX_STOP_MACHINE */
diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c
index 4c89ee9..e8f05b1 100644
--- a/kernel/stop_machine.c
+++ b/kernel/stop_machine.c
@@ -439,8 +439,15 @@ static int stop_machine_cpu_stop(void *data)
 	struct stop_machine_data *smdata = data;
 	enum stopmachine_state curstate = STOPMACHINE_NONE;
 	int cpu = smp_processor_id(), err = 0;
+	unsigned long flags;
 	bool is_active;
 
+	/*
+	 * When called from stop_machine_from_inactive_cpu(), irq might
+	 * already be disabled.  Save the state and restore it on exit.
+	 */
+	local_save_flags(flags);
+
 	if (!smdata->active_cpus)
 		is_active = cpu == cpumask_first(cpu_online_mask);
 	else
@@ -468,7 +475,7 @@ static int stop_machine_cpu_stop(void *data)
 		}
 	} while (curstate != STOPMACHINE_EXIT);
 
-	local_irq_enable();
+	local_irq_restore(flags);
 	return err;
 }
 
@@ -495,4 +502,57 @@ int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
 }
 EXPORT_SYMBOL_GPL(stop_machine);
 
+/**
+ * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
+ * @fn: the function to run
+ * @data: the data ptr for the @fn()
+ * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
+ *
+ * This is identical to stop_machine() but can be called from a CPU which
+ * is not active.  The local CPU is in the process of hotplug (so no other
+ * CPU hotplug can start) and not marked active and doesn't have enough
+ * context to sleep.
+ *
+ * This function provides stop_machine() functionality for such state by
+ * using busy-wait for synchronization and executing @fn directly for local
+ * CPU.
+ *
+ * CONTEXT:
+ * Local CPU is inactive.  Temporarily stops all active CPUs.
+ *
+ * RETURNS:
+ * 0 if all executions of @fn returned 0, any non zero return value if any
+ * returned non zero.
+ */
+int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
+				  const struct cpumask *cpus)
+{
+	struct stop_machine_data smdata = { .fn = fn, .data = data,
+					    .active_cpus = cpus };
+	struct cpu_stop_done done;
+	int ret;
+
+	/* Local CPU must be inactive and CPU hotplug in progress. */
+	BUG_ON(cpu_active(raw_smp_processor_id()));
+	smdata.num_threads = num_active_cpus() + 1;	/* +1 for local */
+
+	/* No proper task established and can't sleep - busy wait for lock. */
+	while (!mutex_trylock(&stop_cpus_mutex))
+		cpu_relax();
+
+	/* Schedule work on other CPUs and execute directly for local CPU */
+	set_state(&smdata, STOPMACHINE_PREPARE);
+	cpu_stop_init_done(&done, num_active_cpus());
+	queue_stop_cpus_work(cpu_active_mask, stop_machine_cpu_stop, &smdata,
+			     &done);
+	ret = stop_machine_cpu_stop(&smdata);
+
+	/* Busy wait for completion. */
+	while (!completion_done(&done.completion))
+		cpu_relax();
+
+	mutex_unlock(&stop_cpus_mutex);
+	return ret ?: done.ret;
+}
+
 #endif	/* CONFIG_STOP_MACHINE */

[tip:x86/mtrr] x86, mtrr: use stop_machine APIs for doing MTRR rendezvous

[tip-bot for Suresh Siddha]

Commit-ID:  192d8857427dd23707d5f0b86ca990c3af6f2d74
Gitweb:     http://git.kernel.org/tip/192d8857427dd23707d5f0b86ca990c3af6f2d74
Author:     Suresh Siddha <suresh.b.siddha@intel.com>
AuthorDate: Thu, 23 Jun 2011 11:19:29 -0700
Committer:  H. Peter Anvin <hpa@linux.intel.com>
CommitDate: Mon, 27 Jun 2011 15:17:13 -0700

x86, mtrr: use stop_machine APIs for doing MTRR rendezvous

MTRR rendezvous sequence is not implemened using stop_machine() before, as this
gets called both from the process context aswell as the cpu online paths
(where the cpu has not come online and the interrupts are disabled etc).

Now that we have a new stop_machine_from_inactive_cpu() API, use it for
rendezvous during mtrr init of a logical processor that is coming online.

For the rest (runtime MTRR modification, system boot, resume paths), use
stop_machine() to implement the rendezvous sequence. This will consolidate and
cleanup the code.

Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Link: http://lkml.kernel.org/r/20110623182057.076997177@sbsiddha-MOBL3.sc.intel.com
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
---
 arch/x86/kernel/cpu/mtrr/main.c |  192 ++++++++------------------------------
 include/linux/stop_machine.h    |    2 -
 kernel/stop_machine.c           |    2 +-
 3 files changed, 42 insertions(+), 154 deletions(-)

diff --git a/arch/x86/kernel/cpu/mtrr/main.c b/arch/x86/kernel/cpu/mtrr/main.c
index 3d17bc7..707b637 100644
--- a/arch/x86/kernel/cpu/mtrr/main.c
+++ b/arch/x86/kernel/cpu/mtrr/main.c
@@ -137,55 +137,43 @@ static void __init init_table(void)
 }
 
 struct set_mtrr_data {
-	atomic_t	count;
-	atomic_t	gate;
 	unsigned long	smp_base;
 	unsigned long	smp_size;
 	unsigned int	smp_reg;
 	mtrr_type	smp_type;
 };
 
-static DEFINE_PER_CPU(struct cpu_stop_work, mtrr_work);
-
 /**
- * mtrr_work_handler - Synchronisation handler. Executed by "other" CPUs.
+ * mtrr_rendezvous_handler - Work done in the synchronization handler. Executed
+ * by all the CPUs.
  * @info: pointer to mtrr configuration data
  *
  * Returns nothing.
  */
-static int mtrr_work_handler(void *info)
+static int mtrr_rendezvous_handler(void *info)
 {
 #ifdef CONFIG_SMP
 	struct set_mtrr_data *data = info;
-	unsigned long flags;
-
-	atomic_dec(&data->count);
-	while (!atomic_read(&data->gate))
-		cpu_relax();
-
-	local_irq_save(flags);
-
-	atomic_dec(&data->count);
-	while (atomic_read(&data->gate))
-		cpu_relax();
 
-	/*  The master has cleared me to execute  */
+	/*
+	 * We use this same function to initialize the mtrrs during boot,
+	 * resume, runtime cpu online and on an explicit request to set a
+	 * specific MTRR.
+	 *
+	 * During boot or suspend, the state of the boot cpu's mtrrs has been
+	 * saved, and we want to replicate that across all the cpus that come
+	 * online (either at the end of boot or resume or during a runtime cpu
+	 * online). If we're doing that, @reg is set to something special and on
+	 * all the cpu's we do mtrr_if->set_all() (On the logical cpu that
+	 * started the boot/resume sequence, this might be a duplicate
+	 * set_all()).
+	 */
 	if (data->smp_reg != ~0U) {
 		mtrr_if->set(data->smp_reg, data->smp_base,
 			     data->smp_size, data->smp_type);
-	} else if (mtrr_aps_delayed_init) {
-		/*
-		 * Initialize the MTRRs inaddition to the synchronisation.
-		 */
+	} else if (mtrr_aps_delayed_init || !cpu_online(smp_processor_id())) {
 		mtrr_if->set_all();
 	}
-
-	atomic_dec(&data->count);
-	while (!atomic_read(&data->gate))
-		cpu_relax();
-
-	atomic_dec(&data->count);
-	local_irq_restore(flags);
 #endif
 	return 0;
 }
@@ -223,20 +211,11 @@ static inline int types_compatible(mtrr_type type1, mtrr_type type2)
  * 14. Wait for buddies to catch up
  * 15. Enable interrupts.
  *
- * What does that mean for us? Well, first we set data.count to the number
- * of CPUs. As each CPU announces that it started the rendezvous handler by
- * decrementing the count, We reset data.count and set the data.gate flag
- * allowing all the cpu's to proceed with the work. As each cpu disables
- * interrupts, it'll decrement data.count once. We wait until it hits 0 and
- * proceed. We clear the data.gate flag and reset data.count. Meanwhile, they
- * are waiting for that flag to be cleared. Once it's cleared, each
- * CPU goes through the transition of updating MTRRs.
- * The CPU vendors may each do it differently,
- * so we call mtrr_if->set() callback and let them take care of it.
- * When they're done, they again decrement data->count and wait for data.gate
- * to be set.
- * When we finish, we wait for data.count to hit 0 and toggle the data.gate flag
- * Everyone then enables interrupts and we all continue on.
+ * What does that mean for us? Well, stop_machine() will ensure that
+ * the rendezvous handler is started on each CPU. And in lockstep they
+ * do the state transition of disabling interrupts, updating MTRR's
+ * (the CPU vendors may each do it differently, so we call mtrr_if->set()
+ * callback and let them take care of it.) and enabling interrupts.
  *
  * Note that the mechanism is the same for UP systems, too; all the SMP stuff
  * becomes nops.
@@ -244,115 +223,26 @@ static inline int types_compatible(mtrr_type type1, mtrr_type type2)
 static void
 set_mtrr(unsigned int reg, unsigned long base, unsigned long size, mtrr_type type)
 {
-	struct set_mtrr_data data;
-	unsigned long flags;
-	int cpu;
-
-#ifdef CONFIG_SMP
-	/*
-	 * If this cpu is not yet active, we are in the cpu online path. There
-	 * can be no stop_machine() in parallel, as stop machine ensures this
-	 * by using get_online_cpus(). We can skip taking the stop_cpus_mutex,
-	 * as we don't need it and also we can't afford to block while waiting
-	 * for the mutex.
-	 *
-	 * If this cpu is active, we need to prevent stop_machine() happening
-	 * in parallel by taking the stop cpus mutex.
-	 *
-	 * Also, this is called in the context of cpu online path or in the
-	 * context where cpu hotplug is prevented. So checking the active status
-	 * of the raw_smp_processor_id() is safe.
-	 */
-	if (cpu_active(raw_smp_processor_id()))
-		mutex_lock(&stop_cpus_mutex);
-#endif
-
-	preempt_disable();
-
-	data.smp_reg = reg;
-	data.smp_base = base;
-	data.smp_size = size;
-	data.smp_type = type;
-	atomic_set(&data.count, num_booting_cpus() - 1);
-
-	/* Make sure data.count is visible before unleashing other CPUs */
-	smp_wmb();
-	atomic_set(&data.gate, 0);
-
-	/* Start the ball rolling on other CPUs */
-	for_each_online_cpu(cpu) {
-		struct cpu_stop_work *work = &per_cpu(mtrr_work, cpu);
-
-		if (cpu == smp_processor_id())
-			continue;
+	struct set_mtrr_data data = { .smp_reg = reg,
+				      .smp_base = base,
+				      .smp_size = size,
+				      .smp_type = type
+				    };
 
-		stop_one_cpu_nowait(cpu, mtrr_work_handler, &data, work);
-	}
-
-
-	while (atomic_read(&data.count))
-		cpu_relax();
-
-	/* Ok, reset count and toggle gate */
-	atomic_set(&data.count, num_booting_cpus() - 1);
-	smp_wmb();
-	atomic_set(&data.gate, 1);
-
-	local_irq_save(flags);
-
-	while (atomic_read(&data.count))
-		cpu_relax();
-
-	/* Ok, reset count and toggle gate */
-	atomic_set(&data.count, num_booting_cpus() - 1);
-	smp_wmb();
-	atomic_set(&data.gate, 0);
-
-	/* Do our MTRR business */
-
-	/*
-	 * HACK!
-	 *
-	 * We use this same function to initialize the mtrrs during boot,
-	 * resume, runtime cpu online and on an explicit request to set a
-	 * specific MTRR.
-	 *
-	 * During boot or suspend, the state of the boot cpu's mtrrs has been
-	 * saved, and we want to replicate that across all the cpus that come
-	 * online (either at the end of boot or resume or during a runtime cpu
-	 * online). If we're doing that, @reg is set to something special and on
-	 * this cpu we still do mtrr_if->set_all(). During boot/resume, this
-	 * is unnecessary if at this point we are still on the cpu that started
-	 * the boot/resume sequence. But there is no guarantee that we are still
-	 * on the same cpu. So we do mtrr_if->set_all() on this cpu aswell to be
-	 * sure that we are in sync with everyone else.
-	 */
-	if (reg != ~0U)
-		mtrr_if->set(reg, base, size, type);
-	else
-		mtrr_if->set_all();
-
-	/* Wait for the others */
-	while (atomic_read(&data.count))
-		cpu_relax();
-
-	atomic_set(&data.count, num_booting_cpus() - 1);
-	smp_wmb();
-	atomic_set(&data.gate, 1);
-
-	/*
-	 * Wait here for everyone to have seen the gate change
-	 * So we're the last ones to touch 'data'
-	 */
-	while (atomic_read(&data.count))
-		cpu_relax();
+	stop_machine(mtrr_rendezvous_handler, &data, cpu_online_mask);
+}
 
-	local_irq_restore(flags);
-	preempt_enable();
-#ifdef CONFIG_SMP
-	if (cpu_active(raw_smp_processor_id()))
-		mutex_unlock(&stop_cpus_mutex);
-#endif
+static void set_mtrr_from_inactive_cpu(unsigned int reg, unsigned long base,
+				      unsigned long size, mtrr_type type)
+{
+	struct set_mtrr_data data = { .smp_reg = reg,
+				      .smp_base = base,
+				      .smp_size = size,
+				      .smp_type = type
+				    };
+
+	stop_machine_from_inactive_cpu(mtrr_rendezvous_handler, &data,
+				       cpu_callout_mask);
 }
 
 /**
@@ -806,7 +696,7 @@ void mtrr_ap_init(void)
 	 *   2. cpu hotadd time. We let mtrr_add/del_page hold cpuhotplug
 	 *      lock to prevent mtrr entry changes
 	 */
-	set_mtrr(~0U, 0, 0, 0);
+	set_mtrr_from_inactive_cpu(~0U, 0, 0, 0);
 }
 
 /**
diff --git a/include/linux/stop_machine.h b/include/linux/stop_machine.h
index e0f2da2..4a9d0c7 100644
--- a/include/linux/stop_machine.h
+++ b/include/linux/stop_machine.h
@@ -27,8 +27,6 @@ struct cpu_stop_work {
 	struct cpu_stop_done	*done;
 };
 
-extern struct mutex stop_cpus_mutex;
-
 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);
diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c
index e8f05b1..c112475 100644
--- a/kernel/stop_machine.c
+++ b/kernel/stop_machine.c
@@ -132,8 +132,8 @@ void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
 	cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), work_buf);
 }
 
-DEFINE_MUTEX(stop_cpus_mutex);
 /* static data for stop_cpus */
+static DEFINE_MUTEX(stop_cpus_mutex);
 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
 
 static void queue_stop_cpus_work(const struct cpumask *cpumask,