Skip to contentSkip to conversation

cpuhp: Expedite RCU when toggling system-wide SMT mode

PATCH Superseded Vishal Chourasia 3 participants 5 messages May 7 – Jun 8

Subsystems (2)

v4 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

v4 cover cpuhp: Expedite RCU when toggling system-wide SMT mode

Hello All,

SMT mode switch operation on a large CPU count system takes close to an hour to complete. Initial debugging root caused the delay to the CPU hotplug subsystem being blocked on numerous synchronize_rcu() calls. Simply enabling system-wide RCU expediting reduced the switch time to 5-6 minutes. Since then, different approaches have been explored, of which some had their own side effects and others didn't work as expected.

Approaches explored:

1. Expedited individual CPU hotplug operations by wrapping _cpu_up()/_cpu_down() with rcu_expedite_gp()/rcu_unexpedite_gp() [0]. Peter suggested expediting only when SMT switch is triggered via the sysfs control interface, not for individual hotplug operations [1].

2. Replacing synchronize_rcu() calls in the CPU hotplug codepath with their expedited variants. This is not viable because one synchronize_rcu() is invoked inside cpus_write_lock(), which is shared with other kernel subsystems [5].

3. Hoisting cpus_write_lock() to be taken once for the entire SMT switch operation instead of per-CPU [3][4]. On large systems where the SMT switch can still take 5-6 minutes, holding the lock for that duration causes hung task splats and starves other subsystems depending on the read lock.

4. Peter has also suggested using rcu_sync_{enter|exit}() which as is doesn't help as is, but can be paired the approach 2 from above.

Current approach: expedite RCU grace periods around the SMT switch operation in the sysfs control interface path, per Peter's suggestion [1], with Aboorva's analysis confirming synchronize_rcu() as the bottleneck [2].

[0] https://lore.kernel.org/all/[email protected]
[1] https://lore.kernel.org/all/[email protected]/
[2] https://lore.kernel.org/all/5f2ab8a44d685701fe36cdaa8042a1aef215d10d.camel@linux.vnet.ibm.com
[3] https://lore.kernel.org/all/[email protected]/
[4] https://lore.kernel.org/all/[email protected]/
[5] https://lore.kernel.org/all/[email protected]/
[6] https://lore.kernel.org/all/[email protected]/

Vishal Chourasia (1):
cpuhp: Expedite RCU when toggling system-wide SMT mode

 include/linux/rcupdate.h | 8 ++++++++
 kernel/cpu.c             | 4 ++++
 kernel/rcu/rcu.h         | 4 ----
 3 files changed, 12 insertions(+), 4 deletions(-)
Signature · 1 line
2.54.0
#1 lore raw

PATCH v4 1/1 cpuhp: Expedite RCU when toggling system-wide SMT mode

On large idle systems, changing system-wide SMT level using the sysfs control interface still takes approximately 40-55 minutes. Changing SMT levels is a user triggered operation generally done when systems are fairly idle. This causes users to be blocked from doing other subsequent work.

Analyzing the collected profile data during SMT level switch, showed that CPU hotplug machinery was blocked on synchronize_rcu() calls.

Expedite RCU grace periods for the entire duration of triggered SMT switch operation via the sysfs control interface. Individual CPU hotplug operations via the online/offline interface are not affected.

On a PPC64 system with 400 CPUs:

SMT8 to SMT1:

    before:  1m14s
    after:   3.2s   (~23x faster)

SMT1 to SMT8:

    before:  2m27s
    after:   2.5s   (~58x faster)

On a large config system with 1920 CPUs, completion time improves from ~1 hour to 5-6 minutes.

---
 include/linux/rcupdate.h | 8 ++++++++
 kernel/cpu.c             | 4 ++++
 kernel/rcu/rcu.h         | 4 ----
 3 files changed, 12 insertions(+), 4 deletions(-)
include/linux/rcupdate.h +8 -0
index bfa765132de8..b6bccf131f1c 100644
@@ -1178,6 +1178,14 @@ rcu_head_after_call_rcu(struct rcu_head *rhp, rcu_callback_t f)
 extern int rcu_expedited;
 extern int rcu_normal;
 
+#ifdef CONFIG_TINY_RCU
+static inline void rcu_expedite_gp(void) { }
+static inline void rcu_unexpedite_gp(void) { }
+#else
+void rcu_expedite_gp(void);
+void rcu_unexpedite_gp(void);
+#endif
+
 DEFINE_LOCK_GUARD_0(rcu, rcu_read_lock(), rcu_read_unlock())
 DECLARE_LOCK_GUARD_0_ATTRS(rcu, __acquires_shared(RCU), __releases_shared(RCU))
 
kernel/cpu.c +4 -0
index bc4f7a9ba64e..6351da9dffdc 100644
@@ -2658,6 +2658,7 @@ int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
 	int cpu, ret = 0;
 
 	cpu_maps_update_begin();
+	rcu_expedite_gp();
 	for_each_online_cpu(cpu) {
 		if (topology_is_primary_thread(cpu))
 			continue;
@@ -2687,6 +2688,7 @@ int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
 	}
 	if (!ret)
 		cpu_smt_control = ctrlval;
+	rcu_unexpedite_gp();
 	cpu_maps_update_done();
 	return ret;
 }
@@ -2704,6 +2706,7 @@ int cpuhp_smt_enable(void)
 	int cpu, ret = 0;
 
 	cpu_maps_update_begin();
+	rcu_expedite_gp();
 	cpu_smt_control = CPU_SMT_ENABLED;
 	for_each_present_cpu(cpu) {
 		/* Skip online CPUs and CPUs on offline nodes */
@@ -2717,6 +2720,7 @@ int cpuhp_smt_enable(void)
 		/* See comment in cpuhp_smt_disable() */
 		cpuhp_online_cpu_device(cpu);
 	}
+	rcu_unexpedite_gp();
 	cpu_maps_update_done();
 	return ret;
 }
kernel/rcu/rcu.h +0 -4
index fa6d30ce73d1..d6a63ee60bc0 100644
@@ -521,8 +521,6 @@ do {									\
 static inline bool rcu_gp_is_normal(void) { return true; }
 static inline bool rcu_gp_is_expedited(void) { return false; }
 static inline bool rcu_async_should_hurry(void) { return false; }
-static inline void rcu_expedite_gp(void) { }
-static inline void rcu_unexpedite_gp(void) { }
 static inline void rcu_async_hurry(void) { }
 static inline void rcu_async_relax(void) { }
 static inline bool rcu_cpu_online(int cpu) { return true; }
@@ -530,8 +528,6 @@ static inline bool rcu_cpu_online(int cpu) { return true; }
 bool rcu_gp_is_normal(void);     /* Internal RCU use. */
 bool rcu_gp_is_expedited(void);  /* Internal RCU use. */
 bool rcu_async_should_hurry(void);  /* Internal RCU use. */
-void rcu_expedite_gp(void);
-void rcu_unexpedite_gp(void);
 void rcu_async_hurry(void);
 void rcu_async_relax(void);
 void rcupdate_announce_bootup_oddness(void);
Signature · 1 line
2.54.0
#2 lore raw
Quoted 23 lines · Vishal Chourasia

On large idle systems, changing system-wide SMT level using the sysfs control interface still takes approximately 40-55 minutes. Changing SMT levels is a user triggered operation generally done when systems are fairly idle. This causes users to be blocked from doing other subsequent work.

Analyzing the collected profile data during SMT level switch, showed that CPU hotplug machinery was blocked on synchronize_rcu() calls.

Expedite RCU grace periods for the entire duration of triggered SMT switch operation via the sysfs control interface. Individual CPU hotplug operations via the online/offline interface are not affected.

On a PPC64 system with 400 CPUs:

   SMT8 to SMT1:
     before:  1m14s
     after:   3.2s   (~23x faster)

   SMT1 to SMT8:
     before:  2m27s
     after:   2.5s   (~58x faster)
Quoted 2 lines

On a large config system with 1920 CPUs, completion time improves from ~1 hour to 5-6 minutes.

Hi Vishal,

I verified the patch on PPC64 system using the configuration described below. Configuration:
    •    Kernel version: 7.1.0-rc2+
    •    Number of CPUs: 960

Using this setup, I evaluated the patch with both SMT enabled and SMT disabled. patch shows a significant improvement in the both SMT=on/off case.

SMT Mode    | Without Patch    | With Patch | % Improvement     |
-------------------------------------------------------------------

SMT=off     | 3m 56.174s       | 0m 34.284s | +85.48%           |
SMT=on      | 3m 47.322s       | 0m 35.583s | +84.35%           |

Regards,
Samir

Quoted 84 lines
---
  include/linux/rcupdate.h | 8 ++++++++
  kernel/cpu.c             | 4 ++++
  kernel/rcu/rcu.h         | 4 ----
  3 files changed, 12 insertions(+), 4 deletions(-)
include/linux/rcupdate.h +8 -0
index bfa765132de8..b6bccf131f1c 100644
@@ -1178,6 +1178,14 @@ rcu_head_after_call_rcu(struct rcu_head *rhp, rcu_callback_t f)
  extern int rcu_expedited;
  extern int rcu_normal;
 
+#ifdef CONFIG_TINY_RCU
+static inline void rcu_expedite_gp(void) { }
+static inline void rcu_unexpedite_gp(void) { }
+#else
+void rcu_expedite_gp(void);
+void rcu_unexpedite_gp(void);
+#endif
+
  DEFINE_LOCK_GUARD_0(rcu, rcu_read_lock(), rcu_read_unlock())
  DECLARE_LOCK_GUARD_0_ATTRS(rcu, __acquires_shared(RCU), __releases_shared(RCU))
 
kernel/cpu.c +4 -0
index bc4f7a9ba64e..6351da9dffdc 100644
@@ -2658,6 +2658,7 @@ int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
  	int cpu, ret = 0;
 
  	cpu_maps_update_begin();
+	rcu_expedite_gp();
  	for_each_online_cpu(cpu) {
  		if (topology_is_primary_thread(cpu))
  			continue;
@@ -2687,6 +2688,7 @@ int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
  	}
  	if (!ret)
  		cpu_smt_control = ctrlval;
+	rcu_unexpedite_gp();
  	cpu_maps_update_done();
  	return ret;
  }
@@ -2704,6 +2706,7 @@ int cpuhp_smt_enable(void)
  	int cpu, ret = 0;
 
  	cpu_maps_update_begin();
+	rcu_expedite_gp();
  	cpu_smt_control = CPU_SMT_ENABLED;
  	for_each_present_cpu(cpu) {
  		/* Skip online CPUs and CPUs on offline nodes */
@@ -2717,6 +2720,7 @@ int cpuhp_smt_enable(void)
  		/* See comment in cpuhp_smt_disable() */
  		cpuhp_online_cpu_device(cpu);
  	}
+	rcu_unexpedite_gp();
  	cpu_maps_update_done();
  	return ret;
  }
kernel/rcu/rcu.h +0 -4
index fa6d30ce73d1..d6a63ee60bc0 100644
@@ -521,8 +521,6 @@ do {									\
  static inline bool rcu_gp_is_normal(void) { return true; }
  static inline bool rcu_gp_is_expedited(void) { return false; }
  static inline bool rcu_async_should_hurry(void) { return false; }
-static inline void rcu_expedite_gp(void) { }
-static inline void rcu_unexpedite_gp(void) { }
  static inline void rcu_async_hurry(void) { }
  static inline void rcu_async_relax(void) { }
  static inline bool rcu_cpu_online(int cpu) { return true; }
@@ -530,8 +528,6 @@ static inline bool rcu_cpu_online(int cpu) { return true; }
  bool rcu_gp_is_normal(void);     /* Internal RCU use. */
  bool rcu_gp_is_expedited(void);  /* Internal RCU use. */
  bool rcu_async_should_hurry(void);  /* Internal RCU use. */
-void rcu_expedite_gp(void);
-void rcu_unexpedite_gp(void);
  void rcu_async_hurry(void);
  void rcu_async_relax(void);
  void rcupdate_announce_bootup_oddness(void);
Reviewed-by on v4 1/1
#3 lore raw
In reply to Vishal Chourasia
Quoted 102 lines · Vishal Chourasia

On large idle systems, changing system-wide SMT level using the sysfs control interface still takes approximately 40-55 minutes. Changing SMT levels is a user triggered operation generally done when systems are fairly idle. This causes users to be blocked from doing other subsequent work.

Analyzing the collected profile data during SMT level switch, showed that CPU hotplug machinery was blocked on synchronize_rcu() calls.

Expedite RCU grace periods for the entire duration of triggered SMT switch operation via the sysfs control interface. Individual CPU hotplug operations via the online/offline interface are not affected.

On a PPC64 system with 400 CPUs:

   SMT8 to SMT1:
     before:  1m14s
     after:   3.2s   (~23x faster)

   SMT1 to SMT8:
     before:  2m27s
     after:   2.5s   (~58x faster)

On a large config system with 1920 CPUs, completion time improves from ~1 hour to 5-6 minutes.

---
  include/linux/rcupdate.h | 8 ++++++++
  kernel/cpu.c             | 4 ++++
  kernel/rcu/rcu.h         | 4 ----
  3 files changed, 12 insertions(+), 4 deletions(-)
include/linux/rcupdate.h +8 -0
index bfa765132de8..b6bccf131f1c 100644
@@ -1178,6 +1178,14 @@ rcu_head_after_call_rcu(struct rcu_head *rhp, rcu_callback_t f)
  extern int rcu_expedited;
  extern int rcu_normal;
 
+#ifdef CONFIG_TINY_RCU
+static inline void rcu_expedite_gp(void) { }
+static inline void rcu_unexpedite_gp(void) { }
+#else
+void rcu_expedite_gp(void);
+void rcu_unexpedite_gp(void);
+#endif
+
  DEFINE_LOCK_GUARD_0(rcu, rcu_read_lock(), rcu_read_unlock())
  DECLARE_LOCK_GUARD_0_ATTRS(rcu, __acquires_shared(RCU), __releases_shared(RCU))
 
kernel/cpu.c +4 -0
index bc4f7a9ba64e..6351da9dffdc 100644
@@ -2658,6 +2658,7 @@ int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
  	int cpu, ret = 0;
 
  	cpu_maps_update_begin();
+	rcu_expedite_gp();
  	for_each_online_cpu(cpu) {
  		if (topology_is_primary_thread(cpu))
  			continue;
@@ -2687,6 +2688,7 @@ int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
  	}
  	if (!ret)
  		cpu_smt_control = ctrlval;
+	rcu_unexpedite_gp();
  	cpu_maps_update_done();
  	return ret;
  }
@@ -2704,6 +2706,7 @@ int cpuhp_smt_enable(void)
  	int cpu, ret = 0;
 
  	cpu_maps_update_begin();
+	rcu_expedite_gp();
  	cpu_smt_control = CPU_SMT_ENABLED;
  	for_each_present_cpu(cpu) {
  		/* Skip online CPUs and CPUs on offline nodes */
@@ -2717,6 +2720,7 @@ int cpuhp_smt_enable(void)
  		/* See comment in cpuhp_smt_disable() */
  		cpuhp_online_cpu_device(cpu);
  	}
+	rcu_unexpedite_gp();
  	cpu_maps_update_done();
  	return ret;
  }
kernel/rcu/rcu.h +0 -2
index fa6d30ce73d1..d6a63ee60bc0 100644
@@ -521,8 +521,6 @@ do {									\
  static inline bool rcu_gp_is_normal(void) { return true; }
  static inline bool rcu_gp_is_expedited(void) { return false; }
  static inline bool rcu_async_should_hurry(void) { return false; }
-static inline void rcu_expedite_gp(void) { }
-static inline void rcu_unexpedite_gp(void) { }
  static inline void rcu_async_hurry(void) { }
  static inline void rcu_async_relax(void) { }
  static inline bool rcu_cpu_online(int cpu) { return true; }

kernel/rcu/rcu.h static inline bool rcu_cpu_online(int cpu) { return true; } · line 533

Quoted 8 lines
kernel/rcu/rcu.h +0 -2
@@ -530,8 +528,6 @@ static inline bool rcu_cpu_online(int cpu) { return true; }
  bool rcu_gp_is_normal(void);     /* Internal RCU use. */
  bool rcu_gp_is_expedited(void);  /* Internal RCU use. */
  bool rcu_async_should_hurry(void);  /* Internal RCU use. */
-void rcu_expedite_gp(void);
-void rcu_unexpedite_gp(void);
  void rcu_async_hurry(void);
  void rcu_async_relax(void);
  void rcupdate_announce_bootup_oddness(void);

Given the context that,
- System is expected to be idle.
- It is user triggered event.

It makes sense to expedite the RCU at the cost of additional IPI's. So

#4 lore raw
In reply to Vishal Chourasia

Hi All,

Gentle ping.
Should I send another version with the tags?

Thanks,
vishalc

Quoted 52 lines · Vishal Chourasia

Hello All,

SMT mode switch operation on a large CPU count system takes close to an hour to complete. Initial debugging root caused the delay to the CPU hotplug subsystem being blocked on numerous synchronize_rcu() calls. Simply enabling system-wide RCU expediting reduced the switch time to 5-6 minutes. Since then, different approaches have been explored, of which some had their own side effects and others didn't work as expected.

Approaches explored:

1. Expedited individual CPU hotplug operations by wrapping _cpu_up()/_cpu_down() with rcu_expedite_gp()/rcu_unexpedite_gp() [0]. Peter suggested expediting only when SMT switch is triggered via the sysfs control interface, not for individual hotplug operations [1].

2. Replacing synchronize_rcu() calls in the CPU hotplug codepath with their expedited variants. This is not viable because one synchronize_rcu() is invoked inside cpus_write_lock(), which is shared with other kernel subsystems [5].

3. Hoisting cpus_write_lock() to be taken once for the entire SMT switch operation instead of per-CPU [3][4]. On large systems where the SMT switch can still take 5-6 minutes, holding the lock for that duration causes hung task splats and starves other subsystems depending on the read lock.

4. Peter has also suggested using rcu_sync_{enter|exit}() which as is doesn't help as is, but can be paired the approach 2 from above.

Current approach: expedite RCU grace periods around the SMT switch operation in the sysfs control interface path, per Peter's suggestion [1], with Aboorva's analysis confirming synchronize_rcu() as the bottleneck [2].

[0] https://lore.kernel.org/all/[email protected]
[1] https://lore.kernel.org/all/[email protected]/
[2] https://lore.kernel.org/all/5f2ab8a44d685701fe36cdaa8042a1aef215d10d.camel@linux.vnet.ibm.com
[3] https://lore.kernel.org/all/[email protected]/
[4] https://lore.kernel.org/all/[email protected]/
[5] https://lore.kernel.org/all/[email protected]/
[6] https://lore.kernel.org/all/[email protected]/

Vishal Chourasia (1):
cpuhp: Expedite RCU when toggling system-wide SMT mode

  include/linux/rcupdate.h | 8 ++++++++
  kernel/cpu.c             | 4 ++++
  kernel/rcu/rcu.h         | 4 ----
  3 files changed, 12 insertions(+), 4 deletions(-)