[PATCH v2 2/3] locking/spinlock_debug: Reduce lock cacheline contention

From: Waiman Long
Date: Wed Feb 01 2017 - 16:14:29 EST


The debug spinlock code is a basic TATAS unfair lock irrespective
of what the underlying architecture specific spinlock implementation
is. As a result, it is sometimes possible to trigger a false positive
"lockup suspected" warning with all the cpu backtraces.

This patch re-implements the debug spinlock as a fair MCS lock. This
reduces the chance of false positive warning messages. At the
same time, it also improves performance by reducing lock cacheline
contention.

Because there is a trylock before entering the MCS queue, this new
debug spinlock code also perform pretty well in a virtual machine
even if its vCPUSs are over-committed.

On a 4-socket 32-core 64-thread system, the performance of a locking
microbenchmark (locking rate and standard deviation) on a 4.9.6 based
debug kernel with and without the patch was as follows:

32 locking threads:

Kernel Locking Rate SD (execution time)
------ ------------ -------------------
w/o patch 263.1 Mop/s 1.39s
with patch 917.6 Mop/s 0.07s

64 locking threads:

Kernel Locking Rate SD (execution time)
------ ------------ -------------------
w/o patch 368.3 Mop/s 6.88s
with patch 733.0 Mop/s 0.09s

On a 2-socket 24-core 48-thread system, the performance of the same
locking microbenchmark (# of locking threads = # of vCPUs) on a KVM
guest are as follows:

24 vCPUs:

Kernel Locking Rate SD (execution time)
------ ------------ -------------------
w/o patch 746.4 Mop/s 1.07s
with patch 1323.6 Mop/s 0.20s

48 vCPUs:

Kernel Locking Rate SD (execution time)
------ ------------ -------------------
w/o patch 1077.8 Mop/s 3.34s
with patch 1090.4 Mop/s 0.29s

72 vCPUs:

Kernel Locking Rate SD (execution time)
------ ------------ -------------------
w/o patch 944.5 Mop/s 3.96s
with patch 1176.7 Mop/s 0.44s

96 vCPUs:

Kernel Locking Rate SD (execution time)
------ ------------ -------------------
w/o patch 878.0 Mop/s 5.19s
with patch 1017.0 Mop/s 0.83s

Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
---
include/linux/spinlock_types.h | 5 ++--
kernel/locking/spinlock_debug.c | 53 +++++++++++++++++++++++++++--------------
2 files changed, 38 insertions(+), 20 deletions(-)

diff --git a/include/linux/spinlock_types.h b/include/linux/spinlock_types.h
index 99f28bd..562af2d 100644
--- a/include/linux/spinlock_types.h
+++ b/include/linux/spinlock_types.h
@@ -25,7 +25,7 @@
#ifdef CONFIG_DEBUG_SPINLOCK
unsigned short lockup;
unsigned int magic, owner_cpu;
- void *owner;
+ void *owner, *tail;
#endif
#ifdef CONFIG_DEBUG_LOCK_ALLOC
struct lockdep_map dep_map;
@@ -47,7 +47,8 @@
.lockup = 0, \
.magic = SPINLOCK_MAGIC, \
.owner_cpu = -1, \
- .owner = SPINLOCK_OWNER_INIT,
+ .owner = SPINLOCK_OWNER_INIT, \
+ .tail = NULL,
#else
# define SPIN_DEBUG_INIT(lockname)
#endif
diff --git a/kernel/locking/spinlock_debug.c b/kernel/locking/spinlock_debug.c
index 0f880a8..c58b61f 100644
--- a/kernel/locking/spinlock_debug.c
+++ b/kernel/locking/spinlock_debug.c
@@ -12,6 +12,7 @@
#include <linux/debug_locks.h>
#include <linux/delay.h>
#include <linux/export.h>
+#include "mcs_spinlock.h"

void __raw_spin_lock_init(raw_spinlock_t *lock, const char *name,
struct lock_class_key *key)
@@ -26,6 +27,7 @@ void __raw_spin_lock_init(raw_spinlock_t *lock, const char *name,
lock->raw_lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
lock->magic = SPINLOCK_MAGIC;
lock->owner = SPINLOCK_OWNER_INIT;
+ lock->tail = NULL;
lock->owner_cpu = -1;
lock->lockup = 0;
}
@@ -105,7 +107,7 @@ static inline void debug_spin_unlock(raw_spinlock_t *lock)
lock->lockup = 0;
}

-static inline void __spin_lockup(raw_spinlock_t *lock)
+static inline void __spin_chk_lockup(raw_spinlock_t *lock, u64 loops)
{
/*
* lockup suspected:
@@ -113,37 +115,52 @@ static inline void __spin_lockup(raw_spinlock_t *lock)
* Only one of the lock waiters will be allowed to print the lockup
* message in order to avoid an avalanche of lockup and backtrace
* messages from different lock waiters of the same lock.
+ *
+ * With the original __deley(1) call, lockup can happen when both
+ * threads of a hyperthreaded CPU core contend on the same lock. So
+ * cpu_relax() is used here instead.
*/
- if (!xchg(&lock->lockup, 1)) {
+ if (unlikely(!loops && !xchg(&lock->lockup, 1))) {
spin_dump(lock, "lockup suspected");
#ifdef CONFIG_SMP
trigger_all_cpu_backtrace();
#endif
}
+ cpu_relax();
}

+/*
+ * The lock waiters are put into a MCS queue to maintain lock fairness
+ * as well as avoiding excessive contention on the lock cacheline. It
+ * also helps to reduce false positive because of unfairness instead
+ * of real lockup.
+ *
+ * The trylock before entering the MCS queue makes this code perform
+ * reasonably well in a virtual machine where some of the lock waiters
+ * may have their vCPUs preempted.
+ */
static void __spin_lock_debug(raw_spinlock_t *lock)
{
- u64 i;
u64 loops = loops_per_jiffy * HZ;
-
- for (i = 0; i < loops; i++) {
- if (arch_spin_trylock(&lock->raw_lock))
- return;
- __delay(1);
+ struct mcs_spinlock node, *prev;
+
+ node.next = NULL;
+ node.locked = 0;
+ prev = xchg(&lock->tail, &node);
+ if (prev) {
+ WRITE_ONCE(prev->next, &node);
+ while (!READ_ONCE(node.locked))
+ __spin_chk_lockup(lock, loops--);
}

- __spin_lockup(lock);
+ while (!arch_spin_trylock(&lock->raw_lock))
+ __spin_chk_lockup(lock, loops--);

- /*
- * The trylock above was causing a livelock. Give the lower level arch
- * specific lock code a chance to acquire the lock. We have already
- * printed a warning/backtrace at this point. The non-debug arch
- * specific code might actually succeed in acquiring the lock. If it is
- * not successful, the end-result is the same - there is no forward
- * progress.
- */
- arch_spin_lock(&lock->raw_lock);
+ if (cmpxchg(&lock->tail, &node, NULL) == &node)
+ return;
+ while (!READ_ONCE(node.next))
+ cpu_relax();
+ WRITE_ONCE(node.next->locked, 1);
}

void do_raw_spin_lock(raw_spinlock_t *lock)
--
1.8.3.1