Re: [PATCH 5/6] locking/rwsem: Unify wait loop

[Waiman Long <longman@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 5211 bytes --]


On 2/23/23 07:26, Peter Zijlstra wrote:
> Now that the reader and writer wait loops are identical, share the
> code.
>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> ---
>   kernel/locking/rwsem.c |  117 +++++++++++++++++++------------------------------
>   1 file changed, 47 insertions(+), 70 deletions(-)
>
> --- a/kernel/locking/rwsem.c
> +++ b/kernel/locking/rwsem.c
> @@ -650,13 +650,11 @@ static void rwsem_mark_wake(struct rw_se
>    * optionally wake up waiters before it returns.
>    */
>   static inline void
> -rwsem_del_wake_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter,
> -		      struct wake_q_head *wake_q)
> +rwsem_del_wake_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
>   		      __releases(&sem->wait_lock)
>   {
>   	bool first = rwsem_first_waiter(sem) == waiter;
> -
> -	wake_q_init(wake_q);
> +	DEFINE_WAKE_Q(wake_q);
>   
>   	/*
>   	 * If the wait_list isn't empty and the waiter to be deleted is
> @@ -664,10 +662,10 @@ rwsem_del_wake_waiter(struct rw_semaphor
>   	 * be eligible to acquire or spin on the lock.
>   	 */
>   	if (rwsem_del_waiter(sem, waiter) && first)
> -		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, wake_q);
> +		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
>   	raw_spin_unlock_irq(&sem->wait_lock);
> -	if (!wake_q_empty(wake_q))
> -		wake_up_q(wake_q);
> +	if (!wake_q_empty(&wake_q))
> +		wake_up_q(&wake_q);
>   }
>   
>   /*
> @@ -993,6 +991,46 @@ static inline void rwsem_cond_wake_waite
>   	rwsem_mark_wake(sem, wake_type, wake_q);
>   }
>   
> +#define waiter_type(_waiter, _r, _w)	\
> +	((_waiter)->type == RWSEM_WAITING_FOR_READ ? (_r) : (_w))
> +
> +static __always_inline struct rw_semaphore *
> +rwsem_waiter_wait(struct rw_semaphore *sem, struct rwsem_waiter *waiter, int state)
> +{
> +	trace_contention_begin(sem, waiter_type(waiter, LCB_F_READ, LCB_F_WRITE));
> +
> +	/* wait to be given the lock */
> +	for (;;) {
> +		set_current_state(state);
> +		if (!smp_load_acquire(&waiter->task)) {
> +			/* Matches rwsem_waiter_wake()'s smp_store_release(). */
> +			break;
> +		}
> +		if (signal_pending_state(state, current)) {
> +			raw_spin_lock_irq(&sem->wait_lock);
> +			if (waiter->task)
> +				goto out_nolock;
> +			raw_spin_unlock_irq(&sem->wait_lock);
> +			/* Ordered by sem->wait_lock against rwsem_mark_wake(). */
> +			break;
> +		}
> +		schedule_preempt_disabled();
> +		lockevent_inc(waiter_type(waiter, rwsem_sleep_reader, rwsem_sleep_writer));
> +	}
> +
> +	__set_current_state(TASK_RUNNING);
> +	lockevent_inc(waiter_type(waiter, rwsem_rlock, rwsem_wlock));
> +	trace_contention_end(sem, 0);
> +	return sem;
> +
> +out_nolock:
> +	rwsem_del_wake_waiter(sem, waiter);
> +	__set_current_state(TASK_RUNNING);

Similar to boqun's comment, we should move __set_current_state() before 
rwsem_del_wake_waiter().

Unfortunately, lockevent_inc() doesn't work with waiter_type() like that 
as the compilation will fail if CONFIG_LOCK_EVENT_COUNTS is enabled.  
Could you include the attached patch in your series and make the 
following changes?

diff --git a/kernel/locking/rwsem.c b/kernel/locking/rwsem.c
index deb0d016a6ce..5b14b0d076fd 100644
--- a/kernel/locking/rwsem.c
+++ b/kernel/locking/rwsem.c
@@ -1021,13 +1021,14 @@ static inline void rwsem_cond_wake_waiter(struct 
rw_semaphore *sem, long count,
         rwsem_mark_wake(sem, wake_type, wake_q);
  }

-#define waiter_type(_waiter, _r, _w)   \
-       ((_waiter)->type == RWSEM_WAITING_FOR_READ ? (_r) : (_w))
+#define waiter_type(_reader, _r, _w)   ((_reader) ? (_r) : (_w))

  static __always_inline struct rw_semaphore *
  rwsem_waiter_wait(struct rw_semaphore *sem, struct rwsem_waiter 
*waiter, int state)
  {
-       trace_contention_begin(sem, waiter_type(waiter, LCB_F_READ, 
LCB_F_WRITE));
+       bool reader = waiter->type == RWSEM_WAITING_FOR_READ;
+
+       trace_contention_begin(sem, waiter_type(reader, LCB_F_READ, 
LCB_F_WRITE));

         /* wait to be given the lock */
         for (;;) {
@@ -1045,18 +1046,18 @@ rwsem_waiter_wait(struct rw_semaphore *sem, 
struct rwsem_waiter *waiter, int sta
                         break;
                 }
                 schedule_preempt_disabled();
-               lockevent_inc(waiter_type(waiter, rwsem_sleep_reader, 
rwsem_sleep_writer));
+               lockevent_cond_inc2(reader, rwsem_sleep_reader, 
rwsem_sleep_writer);
         }

         __set_current_state(TASK_RUNNING);
-       lockevent_inc(waiter_type(waiter, rwsem_rlock, rwsem_wlock));
+       lockevent_cond_inc2(reader, rwsem_rlock, rwsem_wlock);
         trace_contention_end(sem, 0);
         return sem;

  out_nolock:
         rwsem_del_wake_waiter(sem, waiter);
         __set_current_state(TASK_RUNNING);
-       lockevent_inc(waiter_type(waiter, rwsem_rlock_fail, 
rwsem_wlock_fail));
+       lockevent_cond_inc2(reader, rwsem_rlock_fail, rwsem_wlock_fail);
         trace_contention_end(sem, -EINTR);
         return ERR_PTR(-EINTR);
  }

Thanks,
Longman

lockevent_inc

[-- Attachment #2: 0001-locking-lock_events-Add-a-new-lockevent_cond_inc2-he.patch --]
[-- Type: text/x-patch, Size: 1724 bytes --]

From 490fb153006941ec7b576c9e89cb220a0739a95c Mon Sep 17 00:00:00 2001
From: Waiman Long <longman@redhat.com>
Date: Thu, 23 Feb 2023 17:32:06 -0500
Subject: [PATCH] locking/lock_events: Add a new lockevent_cond_inc2() helper

Add a new lockevent_cond_inc2(cond, true_event, false_event) helper
to conditionally increment one of the 2 given events.

Signed-off-by: Waiman Long <longman@redhat.com>
---
 kernel/locking/lock_events.h | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/kernel/locking/lock_events.h b/kernel/locking/lock_events.h
index 8c7e7d25f09c..668c2f1397f6 100644
--- a/kernel/locking/lock_events.h
+++ b/kernel/locking/lock_events.h
@@ -43,6 +43,21 @@ static inline void __lockevent_inc(enum lock_events event, bool cond)
 #define lockevent_inc(ev)	  __lockevent_inc(LOCKEVENT_ ##ev, true)
 #define lockevent_cond_inc(ev, c) __lockevent_inc(LOCKEVENT_ ##ev, c)
 
+/*
+ * Increment either the tevent (cond true) or fevent (cond false)
+ */
+static inline void __lockevent_cond_inc2(bool cond, enum lock_events tevent,
+				         enum lock_events fevent)
+{
+	if (cond)
+		raw_cpu_inc(lockevents[tevent]);
+	else
+		raw_cpu_inc(lockevents[fevent]);
+
+}
+#define lockevent_cond_inc2(c, tev, fev)	\
+	__lockevent_cond_inc2(c, LOCKEVENT_ ##tev, LOCKEVENT_ ##fev)
+
 static inline void __lockevent_add(enum lock_events event, int inc)
 {
 	raw_cpu_add(lockevents[event], inc);
@@ -55,6 +70,7 @@ static inline void __lockevent_add(enum lock_events event, int inc)
 #define lockevent_inc(ev)
 #define lockevent_add(ev, c)
 #define lockevent_cond_inc(ev, c)
+#define lockevent_cond_inc2(c, tev, fev)
 
 #endif /* CONFIG_LOCK_EVENT_COUNTS */
 #endif /* __LOCKING_LOCK_EVENTS_H */
-- 
2.31.1