[patch 17/25] debugobjects: Rework free_object_work()

Thomas Gleixner posted 25 patches 1 month, 3 weeks ago
[patch 17/25] debugobjects: Rework free_object_work()
Posted by Thomas Gleixner 1 month, 3 weeks ago
Convert it to batch processing with intermediate helper functions. This
reduces the final changes for batch processing.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
---
 lib/debugobjects.c |   80 +++++++++++++++++++++++++----------------------------
 1 file changed, 38 insertions(+), 42 deletions(-)

--- a/lib/debugobjects.c
+++ b/lib/debugobjects.c
@@ -35,7 +35,7 @@
  * frequency of 10Hz and about 1024 objects for each freeing operation.
  * So it is freeing at most 10k debug objects per second.
  */
-#define ODEBUG_FREE_WORK_MAX	1024
+#define ODEBUG_FREE_WORK_MAX	(1024 / ODEBUG_BATCH_SIZE)
 #define ODEBUG_FREE_WORK_DELAY	DIV_ROUND_UP(HZ, 10)
 
 struct debug_bucket {
@@ -158,6 +158,21 @@ static bool pool_move_batch(struct obj_p
 	return true;
 }
 
+static bool pool_pop_batch(struct hlist_head *head, struct obj_pool *src)
+{
+	if (!src->cnt)
+		return false;
+
+	for (int i = 0; src->cnt && i < ODEBUG_BATCH_SIZE; i++) {
+		struct hlist_node *node = src->objects.first;
+
+		WRITE_ONCE(src->cnt, src->cnt - 1);
+		hlist_del(node);
+		hlist_add_head(node, head);
+	}
+	return true;
+}
+
 static struct debug_obj *__alloc_object(struct hlist_head *list)
 {
 	struct debug_obj *obj;
@@ -343,55 +358,36 @@ static struct debug_obj *alloc_object(vo
 	return obj;
 }
 
-/*
- * workqueue function to free objects.
- *
- * To reduce contention on the global pool_lock, the actual freeing of
- * debug objects will be delayed if the pool_lock is busy.
- */
+/* workqueue function to free objects. */
 static void free_obj_work(struct work_struct *work)
 {
-	struct debug_obj *obj;
-	unsigned long flags;
-	HLIST_HEAD(tofree);
+	bool free = true;
 
 	WRITE_ONCE(obj_freeing, false);
-	if (!raw_spin_trylock_irqsave(&pool_lock, flags))
+
+	if (!pool_count(&pool_to_free))
 		return;
 
-	if (pool_global.cnt >= pool_global.max_cnt)
-		goto free_objs;
+	for (unsigned int cnt = 0; cnt < ODEBUG_FREE_WORK_MAX; cnt++) {
+		HLIST_HEAD(tofree);
 
-	/*
-	 * The objs on the pool list might be allocated before the work is
-	 * run, so recheck if pool list it full or not, if not fill pool
-	 * list from the global free list. As it is likely that a workload
-	 * may be gearing up to use more and more objects, don't free any
-	 * of them until the next round.
-	 */
-	while (pool_to_free.cnt && pool_global.cnt < pool_global.max_cnt) {
-		obj = hlist_entry(pool_to_free.objects.first, typeof(*obj), node);
-		hlist_del(&obj->node);
-		hlist_add_head(&obj->node, &pool_global.objects);
-		WRITE_ONCE(pool_to_free.cnt, pool_to_free.cnt - 1);
-		WRITE_ONCE(pool_global.cnt, pool_global.cnt + 1);
-	}
-	raw_spin_unlock_irqrestore(&pool_lock, flags);
-	return;
-
-free_objs:
-	/*
-	 * Pool list is already full and there are still objs on the free
-	 * list. Move remaining free objs to a temporary list to free the
-	 * memory outside the pool_lock held region.
-	 */
-	if (pool_to_free.cnt) {
-		hlist_move_list(&pool_to_free.objects, &tofree);
-		WRITE_ONCE(pool_to_free.cnt, 0);
+		/* Acquire and drop the lock for each batch */
+		scoped_guard(raw_spinlock_irqsave, &pool_lock) {
+			if (!pool_to_free.cnt)
+				return;
+
+			/* Refill the global pool if possible */
+			if (pool_move_batch(&pool_global, &pool_to_free)) {
+				/* Don't free as there seems to be demand */
+				free = false;
+			} else if (free) {
+				pool_pop_batch(&tofree, &pool_to_free);
+			} else {
+				return;
+			}
+		}
+		free_object_list(&tofree);
 	}
-	raw_spin_unlock_irqrestore(&pool_lock, flags);
-
-	free_object_list(&tofree);
 }
 
 static void __free_object(struct debug_obj *obj)
Re: [patch 17/25] debugobjects: Rework free_object_work()
Posted by Leizhen (ThunderTown) 1 month, 2 weeks ago

On 2024/10/8 0:50, Thomas Gleixner wrote:
> Convert it to batch processing with intermediate helper functions. This
> reduces the final changes for batch processing.

Reviewed-by: Zhen Lei <thunder.leizhen@huawei.com>

> 
> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
> ---
>  lib/debugobjects.c |   80 +++++++++++++++++++++++++----------------------------
>  1 file changed, 38 insertions(+), 42 deletions(-)
> 
> --- a/lib/debugobjects.c
> +++ b/lib/debugobjects.c
> @@ -35,7 +35,7 @@
>   * frequency of 10Hz and about 1024 objects for each freeing operation.
>   * So it is freeing at most 10k debug objects per second.
>   */
> -#define ODEBUG_FREE_WORK_MAX	1024
> +#define ODEBUG_FREE_WORK_MAX	(1024 / ODEBUG_BATCH_SIZE)
>  #define ODEBUG_FREE_WORK_DELAY	DIV_ROUND_UP(HZ, 10)
>  
>  struct debug_bucket {
> @@ -158,6 +158,21 @@ static bool pool_move_batch(struct obj_p
>  	return true;
>  }
>  
> +static bool pool_pop_batch(struct hlist_head *head, struct obj_pool *src)
> +{
> +	if (!src->cnt)
> +		return false;
> +
> +	for (int i = 0; src->cnt && i < ODEBUG_BATCH_SIZE; i++) {
> +		struct hlist_node *node = src->objects.first;
> +
> +		WRITE_ONCE(src->cnt, src->cnt - 1);
> +		hlist_del(node);
> +		hlist_add_head(node, head);
> +	}
> +	return true;
> +}
> +
>  static struct debug_obj *__alloc_object(struct hlist_head *list)
>  {
>  	struct debug_obj *obj;
> @@ -343,55 +358,36 @@ static struct debug_obj *alloc_object(vo
>  	return obj;
>  }
>  
> -/*
> - * workqueue function to free objects.
> - *
> - * To reduce contention on the global pool_lock, the actual freeing of
> - * debug objects will be delayed if the pool_lock is busy.
> - */
> +/* workqueue function to free objects. */
>  static void free_obj_work(struct work_struct *work)
>  {
> -	struct debug_obj *obj;
> -	unsigned long flags;
> -	HLIST_HEAD(tofree);
> +	bool free = true;
>  
>  	WRITE_ONCE(obj_freeing, false);
> -	if (!raw_spin_trylock_irqsave(&pool_lock, flags))
> +
> +	if (!pool_count(&pool_to_free))
>  		return;
>  
> -	if (pool_global.cnt >= pool_global.max_cnt)
> -		goto free_objs;
> +	for (unsigned int cnt = 0; cnt < ODEBUG_FREE_WORK_MAX; cnt++) {
> +		HLIST_HEAD(tofree);
>  
> -	/*
> -	 * The objs on the pool list might be allocated before the work is
> -	 * run, so recheck if pool list it full or not, if not fill pool
> -	 * list from the global free list. As it is likely that a workload
> -	 * may be gearing up to use more and more objects, don't free any
> -	 * of them until the next round.
> -	 */
> -	while (pool_to_free.cnt && pool_global.cnt < pool_global.max_cnt) {
> -		obj = hlist_entry(pool_to_free.objects.first, typeof(*obj), node);
> -		hlist_del(&obj->node);
> -		hlist_add_head(&obj->node, &pool_global.objects);
> -		WRITE_ONCE(pool_to_free.cnt, pool_to_free.cnt - 1);
> -		WRITE_ONCE(pool_global.cnt, pool_global.cnt + 1);
> -	}
> -	raw_spin_unlock_irqrestore(&pool_lock, flags);
> -	return;
> -
> -free_objs:
> -	/*
> -	 * Pool list is already full and there are still objs on the free
> -	 * list. Move remaining free objs to a temporary list to free the
> -	 * memory outside the pool_lock held region.
> -	 */
> -	if (pool_to_free.cnt) {
> -		hlist_move_list(&pool_to_free.objects, &tofree);
> -		WRITE_ONCE(pool_to_free.cnt, 0);
> +		/* Acquire and drop the lock for each batch */
> +		scoped_guard(raw_spinlock_irqsave, &pool_lock) {
> +			if (!pool_to_free.cnt)
> +				return;
> +
> +			/* Refill the global pool if possible */
> +			if (pool_move_batch(&pool_global, &pool_to_free)) {
> +				/* Don't free as there seems to be demand */
> +				free = false;
> +			} else if (free) {
> +				pool_pop_batch(&tofree, &pool_to_free);
> +			} else {
> +				return;
> +			}
> +		}
> +		free_object_list(&tofree);
>  	}
> -	raw_spin_unlock_irqrestore(&pool_lock, flags);
> -
> -	free_object_list(&tofree);
>  }
>  
>  static void __free_object(struct debug_obj *obj)
> 
> .
> 

-- 
Regards,
  Zhen Lei
[tip: core/debugobjects] debugobjects: Rework free_object_work()
Posted by tip-bot2 for Thomas Gleixner 1 month, 1 week ago
The following commit has been merged into the core/debugobjects branch of tip:

Commit-ID:     9ce99c6d7bfbca71f1e5fa34045ea48cb768f54a
Gitweb:        https://git.kernel.org/tip/9ce99c6d7bfbca71f1e5fa34045ea48cb768f54a
Author:        Thomas Gleixner <tglx@linutronix.de>
AuthorDate:    Mon, 07 Oct 2024 18:50:12 +02:00
Committer:     Thomas Gleixner <tglx@linutronix.de>
CommitterDate: Tue, 15 Oct 2024 17:30:32 +02:00

debugobjects: Rework free_object_work()

Convert it to batch processing with intermediate helper functions. This
reduces the final changes for batch processing.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Zhen Lei <thunder.leizhen@huawei.com>
Link: https://lore.kernel.org/all/20241007164914.015906394@linutronix.de

---
 lib/debugobjects.c | 82 +++++++++++++++++++++------------------------
 1 file changed, 39 insertions(+), 43 deletions(-)

diff --git a/lib/debugobjects.c b/lib/debugobjects.c
index 3700ddf..d5a8538 100644
--- a/lib/debugobjects.c
+++ b/lib/debugobjects.c
@@ -35,7 +35,7 @@
  * frequency of 10Hz and about 1024 objects for each freeing operation.
  * So it is freeing at most 10k debug objects per second.
  */
-#define ODEBUG_FREE_WORK_MAX	1024
+#define ODEBUG_FREE_WORK_MAX	(1024 / ODEBUG_BATCH_SIZE)
 #define ODEBUG_FREE_WORK_DELAY	DIV_ROUND_UP(HZ, 10)
 
 struct debug_bucket {
@@ -158,6 +158,21 @@ static bool pool_move_batch(struct obj_pool *dst, struct obj_pool *src)
 	return true;
 }
 
+static bool pool_pop_batch(struct hlist_head *head, struct obj_pool *src)
+{
+	if (!src->cnt)
+		return false;
+
+	for (int i = 0; src->cnt && i < ODEBUG_BATCH_SIZE; i++) {
+		struct hlist_node *node = src->objects.first;
+
+		WRITE_ONCE(src->cnt, src->cnt - 1);
+		hlist_del(node);
+		hlist_add_head(node, head);
+	}
+	return true;
+}
+
 static struct debug_obj *__alloc_object(struct hlist_head *list)
 {
 	struct debug_obj *obj;
@@ -343,55 +358,36 @@ static struct debug_obj *alloc_object(void *addr, struct debug_bucket *b,
 	return obj;
 }
 
-/*
- * workqueue function to free objects.
- *
- * To reduce contention on the global pool_lock, the actual freeing of
- * debug objects will be delayed if the pool_lock is busy.
- */
+/* workqueue function to free objects. */
 static void free_obj_work(struct work_struct *work)
 {
-	struct debug_obj *obj;
-	unsigned long flags;
-	HLIST_HEAD(tofree);
+	bool free = true;
 
 	WRITE_ONCE(obj_freeing, false);
-	if (!raw_spin_trylock_irqsave(&pool_lock, flags))
-		return;
 
-	if (pool_global.cnt >= pool_global.max_cnt)
-		goto free_objs;
-
-	/*
-	 * The objs on the pool list might be allocated before the work is
-	 * run, so recheck if pool list it full or not, if not fill pool
-	 * list from the global free list. As it is likely that a workload
-	 * may be gearing up to use more and more objects, don't free any
-	 * of them until the next round.
-	 */
-	while (pool_to_free.cnt && pool_global.cnt < pool_global.max_cnt) {
-		obj = hlist_entry(pool_to_free.objects.first, typeof(*obj), node);
-		hlist_del(&obj->node);
-		hlist_add_head(&obj->node, &pool_global.objects);
-		WRITE_ONCE(pool_to_free.cnt, pool_to_free.cnt - 1);
-		WRITE_ONCE(pool_global.cnt, pool_global.cnt + 1);
-	}
-	raw_spin_unlock_irqrestore(&pool_lock, flags);
-	return;
+	if (!pool_count(&pool_to_free))
+		return;
 
-free_objs:
-	/*
-	 * Pool list is already full and there are still objs on the free
-	 * list. Move remaining free objs to a temporary list to free the
-	 * memory outside the pool_lock held region.
-	 */
-	if (pool_to_free.cnt) {
-		hlist_move_list(&pool_to_free.objects, &tofree);
-		WRITE_ONCE(pool_to_free.cnt, 0);
+	for (unsigned int cnt = 0; cnt < ODEBUG_FREE_WORK_MAX; cnt++) {
+		HLIST_HEAD(tofree);
+
+		/* Acquire and drop the lock for each batch */
+		scoped_guard(raw_spinlock_irqsave, &pool_lock) {
+			if (!pool_to_free.cnt)
+				return;
+
+			/* Refill the global pool if possible */
+			if (pool_move_batch(&pool_global, &pool_to_free)) {
+				/* Don't free as there seems to be demand */
+				free = false;
+			} else if (free) {
+				pool_pop_batch(&tofree, &pool_to_free);
+			} else {
+				return;
+			}
+		}
+		free_object_list(&tofree);
 	}
-	raw_spin_unlock_irqrestore(&pool_lock, flags);
-
-	free_object_list(&tofree);
 }
 
 static void __free_object(struct debug_obj *obj)