Skip to contentSkip to conversation

locking/rwsem: Make handoff writer optimistically spin on owner

PATCH-tip Waiman Long 3 participants 4 messages Jun 25, 2019 – Aug 6, 2019 Locking primitives

lore
Share & tools
lore

Waiman Long

author reviewer
#0 lore raw

When the handoff bit is set by a writer, no other tasks other than the setting writer itself is allowed to acquire the lock. If the to-be-handoff'ed writer goes to sleep, there will be a wakeup latency period where the lock is free, but no one can acquire it. That is less than ideal.

To reduce that latency, the handoff writer will now optimistically spin on the owner if it happens to be a on-cpu writer. It will spin until it releases the lock and the to-be-handoff'ed writer can then acquire the lock immediately without any delay. Of course, if the owner is not a on-cpu writer, the to-be-handoff'ed writer will have to sleep anyway.

The optimistic spinning code is also modified to not stop spinning when the handoff bit is set. This will prevent an occasional setting of handoff bit from causing a bunch of optimistic spinners from entering into the wait queue causing significant reduction in throughput.

On a 1-socket 22-core 44-thread Skylake system, the AIM7 shared_memory workload was run with 7000 users. The throughput (jobs/min) of the following kernels were as follows:

1) 5.2-rc6
- 8,092,486
2) 5.2-rc6 + tip's rwsem patches
- 7,567,568
3) 5.2-rc6 + tip's rwsem patches + this patch
- 7,954,545

Using perf-record(1), the %cpu time used by rwsem_down_write_slowpath(), rwsem_down_write_failed() and their callees for the 3 kernels were 1.70%, 5.46% and 2.08% respectively.

---
 kernel/locking/rwsem.c | 32 ++++++++++++++++++++++++++------
 1 file changed, 26 insertions(+), 6 deletions(-)
kernel/locking/rwsem.c +26 -6
index 37524a47f002..5e022fbdb2bb 100644
@@ -720,11 +720,12 @@ rwsem_spin_on_owner(struct rw_semaphore *sem, unsigned long nonspinnable)
 
 	rcu_read_lock();
 	for (;;) {
-		if (atomic_long_read(&sem->count) & RWSEM_FLAG_HANDOFF) {
-			state = OWNER_NONSPINNABLE;
-			break;
-		}
-
+		/*
+		 * When a waiting writer set the handoff flag, it may spin
+		 * on the owner as well. Once that writer acquires the lock,
+		 * we can spin on it. So we don't need to quit even when the
+		 * handoff bit is set.
+		 */
 		new = rwsem_owner_flags(sem, &new_flags);
 		if ((new != owner) || (new_flags != flags)) {
 			state = rwsem_owner_state(new, new_flags, nonspinnable);
@@ -970,6 +971,13 @@ static inline bool rwsem_reader_phase_trylock(struct rw_semaphore *sem,
 {
 	return false;
 }
+
+static inline int
+rwsem_spin_on_owner(struct rw_semaphore *sem, unsigned long nonspinnable)
+{
+	return 0;
+}
+#define OWNER_NULL	1
 #endif
 
 /*
@@ -1190,6 +1198,18 @@ rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
 
 		raw_spin_unlock_irq(&sem->wait_lock);
 
+		/*
+		 * After setting the handoff bit and failing to acquire
+		 * the lock, attempt to spin on owner to accelerate lock
+		 * transfer. If the previous owner is a on-cpu writer and it
+		 * has just released the lock, OWNER_NULL will be returned.
+		 * In this case, we attempt to acquire the lock again
+		 * without sleeping.
+		 */
+		if ((wstate == WRITER_HANDOFF) &&
+		    (rwsem_spin_on_owner(sem, 0) == OWNER_NULL))
+			goto trylock_again;
+
 		/* Block until there are no active lockers. */
 		for (;;) {
 			if (signal_pending_state(state, current))
@@ -1224,7 +1244,7 @@ rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
 				break;
 			}
 		}
-
+trylock_again:
 		raw_spin_lock_irq(&sem->wait_lock);
 	}
 	__set_current_state(TASK_RUNNING);
Signature · 1 line
2.18.1
author reviewer
#1 lore raw
Quoted 93 lines · Waiman Long

When the handoff bit is set by a writer, no other tasks other than the setting writer itself is allowed to acquire the lock. If the to-be-handoff'ed writer goes to sleep, there will be a wakeup latency period where the lock is free, but no one can acquire it. That is less than ideal.

To reduce that latency, the handoff writer will now optimistically spin on the owner if it happens to be a on-cpu writer. It will spin until it releases the lock and the to-be-handoff'ed writer can then acquire the lock immediately without any delay. Of course, if the owner is not a on-cpu writer, the to-be-handoff'ed writer will have to sleep anyway.

The optimistic spinning code is also modified to not stop spinning when the handoff bit is set. This will prevent an occasional setting of handoff bit from causing a bunch of optimistic spinners from entering into the wait queue causing significant reduction in throughput.

On a 1-socket 22-core 44-thread Skylake system, the AIM7 shared_memory workload was run with 7000 users. The throughput (jobs/min) of the following kernels were as follows:

1) 5.2-rc6
- 8,092,486
2) 5.2-rc6 + tip's rwsem patches
- 7,567,568
3) 5.2-rc6 + tip's rwsem patches + this patch
- 7,954,545

Using perf-record(1), the %cpu time used by rwsem_down_write_slowpath(), rwsem_down_write_failed() and their callees for the 3 kernels were 1.70%, 5.46% and 2.08% respectively.

---
 kernel/locking/rwsem.c | 32 ++++++++++++++++++++++++++------
 1 file changed, 26 insertions(+), 6 deletions(-)
kernel/locking/rwsem.c +25 -5
index 37524a47f002..5e022fbdb2bb 100644
@@ -720,11 +720,12 @@ rwsem_spin_on_owner(struct rw_semaphore *sem, unsigned long nonspinnable)
 
 	rcu_read_lock();
 	for (;;) {
-		if (atomic_long_read(&sem->count) & RWSEM_FLAG_HANDOFF) {
-			state = OWNER_NONSPINNABLE;
-			break;
-		}
-
+		/*
+		 * When a waiting writer set the handoff flag, it may spin
+		 * on the owner as well. Once that writer acquires the lock,
+		 * we can spin on it. So we don't need to quit even when the
+		 * handoff bit is set.
+		 */
 		new = rwsem_owner_flags(sem, &new_flags);
 		if ((new != owner) || (new_flags != flags)) {
 			state = rwsem_owner_state(new, new_flags, nonspinnable);
@@ -970,6 +971,13 @@ static inline bool rwsem_reader_phase_trylock(struct rw_semaphore *sem,
 {
 	return false;
 }
+
+static inline int
+rwsem_spin_on_owner(struct rw_semaphore *sem, unsigned long nonspinnable)
+{
+	return 0;
+}
+#define OWNER_NULL	1
 #endif
 
 /*
@@ -1190,6 +1198,18 @@ rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
 
 		raw_spin_unlock_irq(&sem->wait_lock);
 
+		/*
+		 * After setting the handoff bit and failing to acquire
+		 * the lock, attempt to spin on owner to accelerate lock
+		 * transfer. If the previous owner is a on-cpu writer and it
+		 * has just released the lock, OWNER_NULL will be returned.
+		 * In this case, we attempt to acquire the lock again
+		 * without sleeping.
+		 */
+		if ((wstate == WRITER_HANDOFF) &&
+		    (rwsem_spin_on_owner(sem, 0) == OWNER_NULL))
+			goto trylock_again;
+
 		/* Block until there are no active lockers. */
 		for (;;) {
 			if (signal_pending_state(state, current))

kernel/locking/rwsem.c rwsem_down_write_slowpath(struct rw_semaphore *sem, int state) · line 1250

Quoted 8 lines
kernel/locking/rwsem.c +1 -1
@@ -1224,7 +1244,7 @@ rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
 				break;
 			}
 		}
-
+trylock_again:
 		raw_spin_lock_irq(&sem->wait_lock);
 	}
 	__set_current_state(TASK_RUNNING);

Any comment on this patch?

Cheers,
Longman

maintainer
#2 lore raw
In reply to Waiman Long
Quoted 28 lines · Waiman Long

When the handoff bit is set by a writer, no other tasks other than the setting writer itself is allowed to acquire the lock. If the to-be-handoff'ed writer goes to sleep, there will be a wakeup latency period where the lock is free, but no one can acquire it. That is less than ideal.

To reduce that latency, the handoff writer will now optimistically spin on the owner if it happens to be a on-cpu writer. It will spin until it releases the lock and the to-be-handoff'ed writer can then acquire the lock immediately without any delay. Of course, if the owner is not a on-cpu writer, the to-be-handoff'ed writer will have to sleep anyway.

The optimistic spinning code is also modified to not stop spinning when the handoff bit is set. This will prevent an occasional setting of handoff bit from causing a bunch of optimistic spinners from entering into the wait queue causing significant reduction in throughput.

On a 1-socket 22-core 44-thread Skylake system, the AIM7 shared_memory workload was run with 7000 users. The throughput (jobs/min) of the following kernels were as follows:

1) 5.2-rc6
- 8,092,486
2) 5.2-rc6 + tip's rwsem patches
- 7,567,568
3) 5.2-rc6 + tip's rwsem patches + this patch
- 7,954,545

Quoted 5 lines

Using perf-record(1), the %cpu time used by rwsem_down_write_slowpath(), rwsem_down_write_failed() and their callees for the 3 kernels were 1.70%, 5.46% and 2.08% respectively.

Thanks.. sorry for taking so long.

#3 lore raw
In reply to Waiman Long
locking/rwsem: Make handoff writer optimistically spin on owner
Commit-ID:  91d2a812dfb98b3b4dad661529c33bc38d303461
Gitweb:     https://git.kernel.org/tip/91d2a812dfb98b3b4dad661529c33bc38d303461
Author:     Waiman Long <[email protected]>
AuthorDate: Tue, 25 Jun 2019 10:39:13 -0400
Committer:  Peter Zijlstra <[email protected]>
CommitDate: Tue, 6 Aug 2019 12:49:15 +0200

locking/rwsem: Make handoff writer optimistically spin on owner

When the handoff bit is set by a writer, no other tasks other than the setting writer itself is allowed to acquire the lock. If the to-be-handoff'ed writer goes to sleep, there will be a wakeup latency period where the lock is free, but no one can acquire it. That is less than ideal.

To reduce that latency, the handoff writer will now optimistically spin on the owner if it happens to be a on-cpu writer. It will spin until it releases the lock and the to-be-handoff'ed writer can then acquire the lock immediately without any delay. Of course, if the owner is not a on-cpu writer, the to-be-handoff'ed writer will have to sleep anyway.

The optimistic spinning code is also modified to not stop spinning when the handoff bit is set. This will prevent an occasional setting of handoff bit from causing a bunch of optimistic spinners from entering into the wait queue causing significant reduction in throughput.

On a 1-socket 22-core 44-thread Skylake system, the AIM7 shared_memory workload was run with 7000 users. The throughput (jobs/min) of the following kernels were as follows:

1) 5.2-rc6
- 8,092,486
2) 5.2-rc6 + tip's rwsem patches
- 7,567,568
3) 5.2-rc6 + tip's rwsem patches + this patch
- 7,954,545

Using perf-record(1), the %cpu time used by rwsem_down_write_slowpath(), rwsem_down_write_failed() and their callees for the 3 kernels were 1.70%, 5.46% and 2.08% respectively.

Cc Ingo Molnar <[email protected]>
Cc Will Deacon <[email protected]>
Cc huang ying <[email protected]>
Cc Tim Chen <[email protected]>
Cc Linus Torvalds <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Davidlohr Bueso <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
---
 kernel/locking/rwsem.c | 32 ++++++++++++++++++++++++++------
 1 file changed, 26 insertions(+), 6 deletions(-)
kernel/locking/rwsem.c +26 -6
index bd0f0d05724c..354238a08b7a 100644
@@ -724,11 +724,12 @@ rwsem_spin_on_owner(struct rw_semaphore *sem, unsigned long nonspinnable)
 
 	rcu_read_lock();
 	for (;;) {
-		if (atomic_long_read(&sem->count) & RWSEM_FLAG_HANDOFF) {
-			state = OWNER_NONSPINNABLE;
-			break;
-		}
-
+		/*
+		 * When a waiting writer set the handoff flag, it may spin
+		 * on the owner as well. Once that writer acquires the lock,
+		 * we can spin on it. So we don't need to quit even when the
+		 * handoff bit is set.
+		 */
 		new = rwsem_owner_flags(sem, &new_flags);
 		if ((new != owner) || (new_flags != flags)) {
 			state = rwsem_owner_state(new, new_flags, nonspinnable);
@@ -974,6 +975,13 @@ static inline bool rwsem_reader_phase_trylock(struct rw_semaphore *sem,
 {
 	return false;
 }
+
+static inline int
+rwsem_spin_on_owner(struct rw_semaphore *sem, unsigned long nonspinnable)
+{
+	return 0;
+}
+#define OWNER_NULL	1
 #endif
 
 /*
@@ -1206,6 +1214,18 @@ wait:
 
 		raw_spin_unlock_irq(&sem->wait_lock);
 
+		/*
+		 * After setting the handoff bit and failing to acquire
+		 * the lock, attempt to spin on owner to accelerate lock
+		 * transfer. If the previous owner is a on-cpu writer and it
+		 * has just released the lock, OWNER_NULL will be returned.
+		 * In this case, we attempt to acquire the lock again
+		 * without sleeping.
+		 */
+		if ((wstate == WRITER_HANDOFF) &&
+		    (rwsem_spin_on_owner(sem, 0) == OWNER_NULL))
+			goto trylock_again;
+
 		/* Block until there are no active lockers. */
 		for (;;) {
 			if (signal_pending_state(state, current))
@@ -1240,7 +1260,7 @@ wait:
 				break;
 			}
 		}
-
+trylock_again:
 		raw_spin_lock_irq(&sem->wait_lock);
 	}
 	__set_current_state(TASK_RUNNING);