diff mbox

[v4,1/4] lib/percpu-list: Per-cpu list with associated per-cpu locks

Message ID 1456445324-3894-2-git-send-email-Waiman.Long@hpe.com (mailing list archive)
State New, archived
Headers show

Commit Message

Waiman Long Feb. 26, 2016, 12:08 a.m. UTC
Linked list is used everywhere in the Linux kernel. However, if many
threads are trying to add or delete entries into the same linked list,
it can create a performance bottleneck.

This patch introduces a new per-cpu list subystem with associated
per-cpu locks for protecting each of the lists individually. This
allows list entries insertion and deletion operations to happen in
parallel instead of being serialized with a global list and lock.

List entry insertion is strictly per cpu. List deletion, however, can
happen in a cpu other than the one that did the insertion. So we still
need lock to protect the list. Because of that, there may still be
a small amount of contention when deletion is being done.

A new header file include/linux/percpu-list.h will be added with the
associated pcpu_list_head and pcpu_list_node structures. The following
functions are provided to manage the per-cpu list:

 1. int init_pcpu_list_head(struct pcpu_list_head **ppcpu_head)
 2. void pcpu_list_add(struct pcpu_list_node *node,
		       struct pcpu_list_head *head)
 3. void pcpu_list_del(struct pcpu_list *node)

Iteration of all the list entries within a group of per-cpu
lists is done by calling either the pcpu_list_iterate() or
pcpu_list_iterate_safe() functions in a while loop. They correspond
to the list_for_each_entry() and list_for_each_entry_safe() macros
respectively. The iteration states are keep in a pcpu_list_state
structure that is passed to the iteration functions.

Signed-off-by: Waiman Long <Waiman.Long@hpe.com>
---
 include/linux/percpu-list.h |  235 +++++++++++++++++++++++++++++++++++++++++++
 lib/Makefile                |    2 +-
 lib/percpu-list.c           |   97 ++++++++++++++++++
 3 files changed, 333 insertions(+), 1 deletions(-)
 create mode 100644 include/linux/percpu-list.h
 create mode 100644 lib/percpu-list.c

Comments

Jan Kara Feb. 29, 2016, 8:34 a.m. UTC | #1
On Thu 25-02-16 19:08:41, Waiman Long wrote:
> Linked list is used everywhere in the Linux kernel. However, if many
> threads are trying to add or delete entries into the same linked list,
> it can create a performance bottleneck.
> 
> This patch introduces a new per-cpu list subystem with associated
> per-cpu locks for protecting each of the lists individually. This
> allows list entries insertion and deletion operations to happen in
> parallel instead of being serialized with a global list and lock.
> 
> List entry insertion is strictly per cpu. List deletion, however, can
> happen in a cpu other than the one that did the insertion. So we still
> need lock to protect the list. Because of that, there may still be
> a small amount of contention when deletion is being done.
> 
> A new header file include/linux/percpu-list.h will be added with the
> associated pcpu_list_head and pcpu_list_node structures. The following
> functions are provided to manage the per-cpu list:
> 
>  1. int init_pcpu_list_head(struct pcpu_list_head **ppcpu_head)
>  2. void pcpu_list_add(struct pcpu_list_node *node,
> 		       struct pcpu_list_head *head)
>  3. void pcpu_list_del(struct pcpu_list *node)
> 
> Iteration of all the list entries within a group of per-cpu
> lists is done by calling either the pcpu_list_iterate() or
> pcpu_list_iterate_safe() functions in a while loop. They correspond
> to the list_for_each_entry() and list_for_each_entry_safe() macros
> respectively. The iteration states are keep in a pcpu_list_state
> structure that is passed to the iteration functions.
> 
> Signed-off-by: Waiman Long <Waiman.Long@hpe.com>

...

> +static __always_inline bool
> +__pcpu_list_next_cpu(struct pcpu_list_head *head, struct pcpu_list_state *state)
> +{
> +	if (state->lock)
> +		spin_unlock(state->lock);
> +
> +	if (state->cpu++ >= 0)
> +		return false;
> +
> +	if (list_empty(&head->list))
> +		return false;
> +	state->lock = &head->lock;
> +	spin_lock(state->lock);

Thinking about avoiding the overhead for UP case - is there any point in
spin_lock in UP? You could just do preempt_disable(), couldn't you? And
then you don't have to pass around any spinlock pointers, don't you?

> +	state->curr = list_entry(head->list.next, struct pcpu_list_node, list);
> +	return true;
> +

^^^ Spurious empty line

> +}
> +#else /* NR_CPUS == 1 */
> +/*
> + * Multiprocessor
> + */
> +static inline bool pcpu_list_empty(struct pcpu_list_head *pcpu_head)
> +{
> +	int cpu;
> +
> +	for_each_possible_cpu(cpu)
> +		if (!list_empty(&per_cpu_ptr(pcpu_head, cpu)->list))
> +			return false;
> +	return true;
> +}
> +
> +/*
> + * Helper function to find the first entry of the next per-cpu list
> + * It works somewhat like for_each_possible_cpu(cpu).
> + *
> + * Return: true if the entry is found, false if all the lists exhausted
> + */
> +static __always_inline bool
> +__pcpu_list_next_cpu(struct pcpu_list_head *head, struct pcpu_list_state *state)
> +{
> +	if (state->lock)
> +		spin_unlock(state->lock);
> +next_cpu:
> +	/*
> +	 * for_each_possible_cpu(cpu)
> +	 */
> +	state->cpu = cpumask_next(state->cpu, cpu_possible_mask);
> +	if (state->cpu >= nr_cpu_ids)
> +		return false;	/* All the per-cpu lists iterated */
> +
> +	state->head = &per_cpu_ptr(head, state->cpu)->list;
> +	if (list_empty(state->head))
> +		goto next_cpu;
> +
> +	state->lock = &per_cpu_ptr(head, state->cpu)->lock;
> +	spin_lock(state->lock);
> +	state->curr = list_entry(state->head->next,
> +				 struct pcpu_list_node, list);

And I've realized this is still racy. The list can be empty by the time you
acquire state->lock so you have to recheck list_empty() after acquiring the
lock.

								Honza
diff mbox

Patch

diff --git a/include/linux/percpu-list.h b/include/linux/percpu-list.h
new file mode 100644
index 0000000..0bb894a
--- /dev/null
+++ b/include/linux/percpu-list.h
@@ -0,0 +1,235 @@ 
+/*
+ * Per-cpu list
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * (C) Copyright 2016 Hewlett-Packard Enterprise Development LP
+ *
+ * Authors: Waiman Long <waiman.long@hpe.com>
+ */
+#ifndef __LINUX_PERCPU_LIST_H
+#define __LINUX_PERCPU_LIST_H
+
+#include <linux/spinlock.h>
+#include <linux/list.h>
+#include <linux/percpu.h>
+
+/*
+ * include/linux/percpu-list.h
+ *
+ * A per-cpu list protected by a per-cpu spinlock.
+ *
+ * The pcpu_list_head structure contains the spinlock, the other
+ * pcpu_list_node structures only contains a pointer to the spinlock in
+ * pcpu_list_head.
+ */
+struct pcpu_list_head {
+	struct list_head list;
+	spinlock_t lock;
+};
+
+struct pcpu_list_node {
+	struct list_head list;
+	spinlock_t *lockptr;
+};
+
+/*
+ * Per-cpu list iteration state
+ */
+struct pcpu_list_state {
+	int			 cpu;
+	spinlock_t		*lock;
+	struct list_head	*head;	/* List head of current per-cpu list */
+	struct pcpu_list_node	*curr;
+	struct pcpu_list_node	*next;
+};
+
+#define PCPU_LIST_HEAD_INIT(name)				\
+	{							\
+		.list.prev = &name.list,			\
+		.list.next = &name.list,			\
+		.list.lock = __SPIN_LOCK_UNLOCKED(name),	\
+	}
+
+#define PCPU_LIST_NODE_INIT(name)		\
+	{					\
+		.list.prev = &name.list,	\
+		.list.next = &name.list,	\
+		.list.lockptr = NULL		\
+	}
+
+#define PCPU_LIST_STATE_INIT()			\
+	{					\
+		.cpu  = -1,			\
+		.lock = NULL,			\
+		.head = NULL,			\
+		.curr = NULL,			\
+		.next = NULL,			\
+	}
+
+#define DEFINE_PCPU_LIST_STATE(s)		\
+	struct pcpu_list_state s = PCPU_LIST_STATE_INIT()
+
+#define pcpu_list_next_entry(pos, member) list_next_entry(pos, member.list)
+
+static inline void init_pcpu_list_node(struct pcpu_list_node *node)
+{
+	INIT_LIST_HEAD(&node->list);
+	node->lockptr = NULL;
+}
+
+static inline void free_pcpu_list_head(struct pcpu_list_head **ppcpu_head)
+{
+	free_percpu(*ppcpu_head);
+	*ppcpu_head = NULL;
+}
+
+static inline void init_pcpu_list_state(struct pcpu_list_state *state)
+{
+	state->cpu  = -1;
+	state->lock = NULL;
+	state->head = NULL;
+	state->curr = NULL;
+	state->next = NULL;
+}
+
+#if NR_CPUS == 1
+/*
+ * For uniprocessor, the list head and lock in struct pcpu_list_head are
+ * used directly.
+ */
+static inline bool pcpu_list_empty(struct pcpu_list_head *pcpu_head)
+{
+	return list_empty(&pcpu_head->list);
+}
+
+static __always_inline bool
+__pcpu_list_next_cpu(struct pcpu_list_head *head, struct pcpu_list_state *state)
+{
+	if (state->lock)
+		spin_unlock(state->lock);
+
+	if (state->cpu++ >= 0)
+		return false;
+
+	if (list_empty(&head->list))
+		return false;
+	state->lock = &head->lock;
+	spin_lock(state->lock);
+	state->curr = list_entry(head->list.next, struct pcpu_list_node, list);
+	return true;
+
+}
+#else /* NR_CPUS == 1 */
+/*
+ * Multiprocessor
+ */
+static inline bool pcpu_list_empty(struct pcpu_list_head *pcpu_head)
+{
+	int cpu;
+
+	for_each_possible_cpu(cpu)
+		if (!list_empty(&per_cpu_ptr(pcpu_head, cpu)->list))
+			return false;
+	return true;
+}
+
+/*
+ * Helper function to find the first entry of the next per-cpu list
+ * It works somewhat like for_each_possible_cpu(cpu).
+ *
+ * Return: true if the entry is found, false if all the lists exhausted
+ */
+static __always_inline bool
+__pcpu_list_next_cpu(struct pcpu_list_head *head, struct pcpu_list_state *state)
+{
+	if (state->lock)
+		spin_unlock(state->lock);
+next_cpu:
+	/*
+	 * for_each_possible_cpu(cpu)
+	 */
+	state->cpu = cpumask_next(state->cpu, cpu_possible_mask);
+	if (state->cpu >= nr_cpu_ids)
+		return false;	/* All the per-cpu lists iterated */
+
+	state->head = &per_cpu_ptr(head, state->cpu)->list;
+	if (list_empty(state->head))
+		goto next_cpu;
+
+	state->lock = &per_cpu_ptr(head, state->cpu)->lock;
+	spin_lock(state->lock);
+	state->curr = list_entry(state->head->next,
+				 struct pcpu_list_node, list);
+	return true;
+}
+#endif /* NR_CPUS == 1 */
+
+/*
+ * Iterate to the next entry of the group of per-cpu lists
+ *
+ * Return: true if the next entry is found, false if all the entries iterated
+ */
+static inline bool pcpu_list_iterate(struct pcpu_list_head *head,
+				     struct pcpu_list_state *state)
+{
+	/*
+	 * Find next entry
+	 */
+	if (state->curr)
+		state->curr = list_next_entry(state->curr, list);
+
+	if (!state->curr || (&state->curr->list == state->head)) {
+		/*
+		 * The current per-cpu list has been exhausted, try the next
+		 * per-cpu list.
+		 */
+		if (!__pcpu_list_next_cpu(head, state))
+			return false;
+	}
+	return true;	/* Continue the iteration */
+}
+
+/*
+ * Iterate to the next entry of the group of per-cpu lists and safe
+ * against removal of list_entry
+ *
+ * Return: true if the next entry is found, false if all the entries iterated
+ */
+static inline bool pcpu_list_iterate_safe(struct pcpu_list_head *head,
+					  struct pcpu_list_state *state)
+{
+	/*
+	 * Find next entry
+	 */
+	if (state->curr) {
+		state->curr = state->next;
+		state->next = list_next_entry(state->next, list);
+	}
+
+	if (!state->curr || (&state->curr->list == state->head)) {
+		/*
+		 * The current per-cpu list has been exhausted, try the next
+		 * per-cpu list.
+		 */
+		if (!__pcpu_list_next_cpu(head, state))
+			return false;
+		state->next = list_next_entry(state->curr, list);
+	}
+	return true;	/* Continue the iteration */
+}
+
+extern int  init_pcpu_list_head(struct pcpu_list_head **ppcpu_head);
+extern void pcpu_list_add(struct pcpu_list_node *node,
+			  struct pcpu_list_head *head);
+extern void pcpu_list_del(struct pcpu_list_node *node);
+
+#endif /* __LINUX_PERCPU_LIST_H */
diff --git a/lib/Makefile b/lib/Makefile
index a7c26a4..71a25d4 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -27,7 +27,7 @@  obj-y += bcd.o div64.o sort.o parser.o halfmd4.o debug_locks.o random32.o \
 	 gcd.o lcm.o list_sort.o uuid.o flex_array.o iov_iter.o clz_ctz.o \
 	 bsearch.o find_bit.o llist.o memweight.o kfifo.o \
 	 percpu-refcount.o percpu_ida.o rhashtable.o reciprocal_div.o \
-	 once.o
+	 once.o percpu-list.o
 obj-y += string_helpers.o
 obj-$(CONFIG_TEST_STRING_HELPERS) += test-string_helpers.o
 obj-y += hexdump.o
diff --git a/lib/percpu-list.c b/lib/percpu-list.c
new file mode 100644
index 0000000..c2d8138
--- /dev/null
+++ b/lib/percpu-list.c
@@ -0,0 +1,97 @@ 
+/*
+ * Per-cpu list
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * (C) Copyright 2016 Hewlett-Packard Enterprise Development LP
+ *
+ * Authors: Waiman Long <waiman.long@hpe.com>
+ */
+#include <linux/percpu-list.h>
+#include <linux/lockdep.h>
+
+/*
+ * The per-cpu list lock needs its own class to avoid warning and stack
+ * trace when lockdep is enabled.
+ */
+static struct lock_class_key percpu_list_key;
+
+/*
+ * Initialize the per-cpu list
+ */
+int init_pcpu_list_head(struct pcpu_list_head **ppcpu_head)
+{
+	struct pcpu_list_head *pcpu_head = alloc_percpu(struct pcpu_list_head);
+	int cpu;
+
+	if (!pcpu_head)
+		return -ENOMEM;
+
+	for_each_possible_cpu(cpu) {
+		struct pcpu_list_head *head = per_cpu_ptr(pcpu_head, cpu);
+
+		INIT_LIST_HEAD(&head->list);
+		head->lock = __SPIN_LOCK_UNLOCKED(&head->lock);
+		lockdep_set_class(&head->lock, &percpu_list_key);
+	}
+
+	*ppcpu_head = pcpu_head;
+	return 0;
+}
+
+/*
+ * List selection is based on the CPU being used when the pcpu_list_add()
+ * function is called. However, deletion may be done by a different CPU.
+ * So we still need to use a lock to protect the content of the list.
+ */
+void pcpu_list_add(struct pcpu_list_node *node, struct pcpu_list_head *head)
+{
+	struct pcpu_list_head *myhead = this_cpu_ptr(head);
+
+	/*
+	 * There is a very slight chance the cpu will be changed
+	 * (by preemption) before calling spin_lock(). We only need to put
+	 * the node in one of the per-cpu lists. It may not need to be
+	 * that of the current cpu.
+	 */
+	spin_lock(&myhead->lock);
+	node->lockptr = &myhead->lock;
+	list_add(&node->list, &myhead->list);
+	spin_unlock(&myhead->lock);
+}
+
+/*
+ * Delete a node from a percpu list
+ *
+ * We need to check the lock pointer again after taking the lock to guard
+ * against concurrent delete of the same node. If the lock pointer changes
+ * (becomes NULL or to a different one), we assume that the deletion was done
+ * elsewhere.
+ */
+void pcpu_list_del(struct pcpu_list_node *node)
+{
+	spinlock_t *lock = READ_ONCE(node->lockptr);
+
+	if (unlikely(!lock))
+		return;
+
+	spin_lock(lock);
+	if (likely(lock == node->lockptr)) {
+		list_del_init(&node->list);
+		node->lockptr = NULL;
+	} else {
+		/*
+		 * This path should never be executed.
+		 */
+		WARN_ON(1);
+	}
+	spin_unlock(lock);
+}