[PATCH -tip 0/2] locking/rwbase: Two reader optimizations

[PATCH -tip 0/2] locking/rwbase: Two reader optimizations

[Davidlohr Bueso]

Hi,

Patch 1 is a barrier optimization that came up from the reader
fastpath ordering auditing.

Patch 2 is a resend of the previous broken patch that attempts
to use wake_q for read_unlock() slowpath.

Tested on v5.15.y-rt. Applies against tip/urgent.

Thanks!

Davidlohr Bueso (2):
  locking/rwbase: Optimize rwbase_read_trylock
  locking/rwbase: Lockless reader waking up a writer

 kernel/locking/rtmutex.c   | 19 +++++++++++++------
 kernel/locking/rwbase_rt.c | 11 +++++++----
 2 files changed, 20 insertions(+), 10 deletions(-)

--
2.26.2

[PATCH 1/2] locking/rwbase: Optimize rwbase_read_trylock

[Davidlohr Bueso]

Instead of a full barrier around the Rmw insn, micro-optimize
for weakly ordered archs such that we only provide the required
ACQUIRE semantics when taking the read lock.

Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
---
 kernel/locking/rwbase_rt.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/kernel/locking/rwbase_rt.c b/kernel/locking/rwbase_rt.c
index 88191f6e252c..a9034784a5a0 100644
--- a/kernel/locking/rwbase_rt.c
+++ b/kernel/locking/rwbase_rt.c
@@ -59,8 +59,7 @@ static __always_inline int rwbase_read_trylock(struct rwbase_rt *rwb)
 	 * set.
 	 */
 	for (r = atomic_read(&rwb->readers); r < 0;) {
-		/* Fully-ordered if cmpxchg() succeeds, provides ACQUIRE */
-		if (likely(atomic_try_cmpxchg(&rwb->readers, &r, r + 1)))
+		if (likely(atomic_try_cmpxchg_acquire(&rwb->readers, &r, r + 1)))
 			return 1;
 	}
 	return 0;
@@ -183,7 +182,7 @@ static inline void __rwbase_write_unlock(struct rwbase_rt *rwb, int bias,
 
 	/*
 	 * _release() is needed in case that reader is in fast path, pairing
-	 * with atomic_try_cmpxchg() in rwbase_read_trylock(), provides RELEASE
+	 * with atomic_try_cmpxchg_acquire() in rwbase_read_trylock().
 	 */
 	(void)atomic_add_return_release(READER_BIAS - bias, &rwb->readers);
 	raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
-- 
2.26.2

[PATCH 2/2] locking/rwbase: Lockless reader waking up a writer

[Davidlohr Bueso]

Use the RT-lock safe wake_q to allow waking up the writer
without having to hold the wait_lock across the operation.

While this is ideally for batching wakeups, single wakeup
usage has still shown to be beneficial vs the cost of
try_to_wakeup() when the lock is contended, as well as
not having irqs disabled during the wakeup window, albeit
preemption will remain disabled.

Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
---
 kernel/locking/rtmutex.c   | 19 +++++++++++++------
 kernel/locking/rwbase_rt.c |  6 +++++-
 2 files changed, 18 insertions(+), 7 deletions(-)

diff --git a/kernel/locking/rtmutex.c b/kernel/locking/rtmutex.c
index 6bb116c559b4..1581674d640b 100644
--- a/kernel/locking/rtmutex.c
+++ b/kernel/locking/rtmutex.c
@@ -446,19 +446,26 @@ static __always_inline void rt_mutex_adjust_prio(struct task_struct *p)
 }
 
 /* RT mutex specific wake_q wrappers */
-static __always_inline void rt_mutex_wake_q_add(struct rt_wake_q_head *wqh,
-						struct rt_mutex_waiter *w)
+static __always_inline void rt_mutex_wake_q_add_task(struct rt_wake_q_head *wqh,
+						     struct task_struct *task,
+						     unsigned int wake_state)
 {
-	if (IS_ENABLED(CONFIG_PREEMPT_RT) && w->wake_state != TASK_NORMAL) {
+	if (IS_ENABLED(CONFIG_PREEMPT_RT) && wake_state != TASK_NORMAL) {
 		if (IS_ENABLED(CONFIG_PROVE_LOCKING))
 			WARN_ON_ONCE(wqh->rtlock_task);
-		get_task_struct(w->task);
-		wqh->rtlock_task = w->task;
+		get_task_struct(task);
+		wqh->rtlock_task = task;
 	} else {
-		wake_q_add(&wqh->head, w->task);
+		wake_q_add(&wqh->head, task);
 	}
 }
 
+static __always_inline void rt_mutex_wake_q_add(struct rt_wake_q_head *wqh,
+						struct rt_mutex_waiter *w)
+{
+	rt_mutex_wake_q_add_task(wqh, w->task, w->wake_state);
+}
+
 static __always_inline void rt_mutex_wake_up_q(struct rt_wake_q_head *wqh)
 {
 	if (IS_ENABLED(CONFIG_PREEMPT_RT) && wqh->rtlock_task) {
diff --git a/kernel/locking/rwbase_rt.c b/kernel/locking/rwbase_rt.c
index a9034784a5a0..8cb58758af3d 100644
--- a/kernel/locking/rwbase_rt.c
+++ b/kernel/locking/rwbase_rt.c
@@ -147,6 +147,7 @@ static void __sched __rwbase_read_unlock(struct rwbase_rt *rwb,
 {
 	struct rt_mutex_base *rtm = &rwb->rtmutex;
 	struct task_struct *owner;
+	DEFINE_RT_WAKE_Q(wqh);
 
 	raw_spin_lock_irq(&rtm->wait_lock);
 	/*
@@ -157,9 +158,12 @@ static void __sched __rwbase_read_unlock(struct rwbase_rt *rwb,
 	 */
 	owner = rt_mutex_owner(rtm);
 	if (owner)
-		wake_up_state(owner, state);
+		rt_mutex_wake_q_add_task(&wqh, owner, state);
 
+	/* Pairs with the preempt_enable() in rt_mutex_wake_up_q() */
+	preempt_disable();
 	raw_spin_unlock_irq(&rtm->wait_lock);
+	rt_mutex_wake_up_q(&wqh);
 }
 
 static __always_inline void rwbase_read_unlock(struct rwbase_rt *rwb,
-- 
2.26.2

Re: [PATCH -tip 0/2] locking/rwbase: Two reader optimizations

[Waiman Long]

On 9/20/21 1:20 AM, Davidlohr Bueso wrote:
> Hi,
>
> Patch 1 is a barrier optimization that came up from the reader
> fastpath ordering auditing.
>
> Patch 2 is a resend of the previous broken patch that attempts
> to use wake_q for read_unlock() slowpath.
>
> Tested on v5.15.y-rt. Applies against tip/urgent.
>
> Thanks!
>
> Davidlohr Bueso (2):
>    locking/rwbase: Optimize rwbase_read_trylock
>    locking/rwbase: Lockless reader waking up a writer
>
>   kernel/locking/rtmutex.c   | 19 +++++++++++++------
>   kernel/locking/rwbase_rt.c | 11 +++++++----
>   2 files changed, 20 insertions(+), 10 deletions(-)
>
> --
> 2.26.2
>
Your patches look good to me.

Acked-by: Waiman Long <longman@redhat.com>

[tip: locking/core] locking/rwbase: Optimize rwbase_read_trylock

[tip-bot2 for Davidlohr Bueso]

The following commit has been merged into the locking/core branch of tip:

Commit-ID:     c78416d122243c92992a1d1063f17ddd0bc80e6c
Gitweb:        https://git.kernel.org/tip/c78416d122243c92992a1d1063f17ddd0bc80e6c
Author:        Davidlohr Bueso <dave@stgolabs.net>
AuthorDate:    Sun, 19 Sep 2021 22:20:30 -07:00
Committer:     Peter Zijlstra <peterz@infradead.org>
CommitterDate: Thu, 07 Oct 2021 13:51:07 +02:00

locking/rwbase: Optimize rwbase_read_trylock

Instead of a full barrier around the Rmw insn, micro-optimize
for weakly ordered archs such that we only provide the required
ACQUIRE semantics when taking the read lock.

Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Waiman Long <longman@redhat.com>
Link: https://lkml.kernel.org/r/20210920052031.54220-2-dave@stgolabs.net
---
 kernel/locking/rwbase_rt.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/kernel/locking/rwbase_rt.c b/kernel/locking/rwbase_rt.c
index 15c8110..6fd3162 100644
--- a/kernel/locking/rwbase_rt.c
+++ b/kernel/locking/rwbase_rt.c
@@ -59,8 +59,7 @@ static __always_inline int rwbase_read_trylock(struct rwbase_rt *rwb)
 	 * set.
 	 */
 	for (r = atomic_read(&rwb->readers); r < 0;) {
-		/* Fully-ordered if cmpxchg() succeeds, provides ACQUIRE */
-		if (likely(atomic_try_cmpxchg(&rwb->readers, &r, r + 1)))
+		if (likely(atomic_try_cmpxchg_acquire(&rwb->readers, &r, r + 1)))
 			return 1;
 	}
 	return 0;
@@ -187,7 +186,7 @@ static inline void __rwbase_write_unlock(struct rwbase_rt *rwb, int bias,
 
 	/*
 	 * _release() is needed in case that reader is in fast path, pairing
-	 * with atomic_try_cmpxchg() in rwbase_read_trylock(), provides RELEASE
+	 * with atomic_try_cmpxchg_acquire() in rwbase_read_trylock().
 	 */
 	(void)atomic_add_return_release(READER_BIAS - bias, &rwb->readers);
 	raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);