Date:	Tue, 2 Jul 2013 14:23:51 -0700
From:	ecashin@...erose.net
To:	Andrew Morton <akpm@...ux-foundation.org>
Cc:	linux-kernel@...r.kernel.org, ecashin@...aid.com
Subject: [PATCH 1/3] aoe: perform I/O completions in parallel

Some users have a large AoE target while others like to use
many AoE targets at the same time.  In the latter case, there
is an opportunity to greatly improve aggregate throughput by
allowing different threads to complete the I/O associated
with each target.  For 36 targets, 4 KiB read throughput
roughly doubles, for example, with these changes in place.

Signed-off-by: Ed Cashin <ecashin@...aid.com>
---
 drivers/block/aoe/aoe.h    |    7 ++-
 drivers/block/aoe/aoecmd.c |  153 ++++++++++++++++++++++++++++++++++++--------
 drivers/block/aoe/aoedev.c |    1 -
 drivers/block/aoe/aoenet.c |    5 +-
 4 files changed, 135 insertions(+), 31 deletions(-)

diff --git a/drivers/block/aoe/aoe.h b/drivers/block/aoe/aoe.h
index 1756494..c969852 100644
--- a/drivers/block/aoe/aoe.h
+++ b/drivers/block/aoe/aoe.h
@@ -196,9 +196,11 @@ struct ktstate {
 	struct completion rendez;
 	struct task_struct *task;
 	wait_queue_head_t *waitq;
-	int (*fn) (void);
-	char *name;
+	int (*fn) (int);
+	char name[12];
 	spinlock_t *lock;
+	int id;
+	int active;
 };
 
 int aoeblk_init(void);
@@ -222,6 +224,7 @@ int aoecmd_init(void);
 struct sk_buff *aoecmd_ata_id(struct aoedev *);
 void aoe_freetframe(struct frame *);
 void aoe_flush_iocq(void);
+void aoe_flush_iocq_by_index(int);
 void aoe_end_request(struct aoedev *, struct request *, int);
 int aoe_ktstart(struct ktstate *k);
 void aoe_ktstop(struct ktstate *k);
diff --git a/drivers/block/aoe/aoecmd.c b/drivers/block/aoe/aoecmd.c
index b75c7db..2ab5455 100644
--- a/drivers/block/aoe/aoecmd.c
+++ b/drivers/block/aoe/aoecmd.c
@@ -35,14 +35,24 @@ module_param(aoe_maxout, int, 0644);
 MODULE_PARM_DESC(aoe_maxout,
 	"Only aoe_maxout outstanding packets for every MAC on eX.Y.");
 
-static wait_queue_head_t ktiowq;
-static struct ktstate kts;
+/* The number of online cpus in the system,
+ * this dictates the number of ktio threads
+ * which will be spawned.
+ */
+static int ncpus;
+
+/* mutex lock used for synchronization while thread spawning */
+static DEFINE_MUTEX(ktio_spawn_lock);
+
+static wait_queue_head_t *ktiowq;
+static struct ktstate *kts;
 
 /* io completion queue */
-static struct {
+struct iocq_ktio {
 	struct list_head head;
 	spinlock_t lock;
-} iocq;
+};
+static struct iocq_ktio *iocq;
 
 static struct page *empty_page;
 
@@ -1278,23 +1288,36 @@ out:
  * Returns true iff responses needing processing remain.
  */
 static int
-ktio(void)
+ktio(int id)
 {
 	struct frame *f;
 	struct list_head *pos;
 	int i;
+	int actual_id;
 
 	for (i = 0; ; ++i) {
 		if (i == MAXIOC)
 			return 1;
-		if (list_empty(&iocq.head))
+		if (list_empty(&iocq[id].head))
 			return 0;
-		pos = iocq.head.next;
+		pos = iocq[id].head.next;
 		list_del(pos);
-		spin_unlock_irq(&iocq.lock);
 		f = list_entry(pos, struct frame, head);
+		spin_unlock_irq(&iocq[id].lock);
 		ktiocomplete(f);
-		spin_lock_irq(&iocq.lock);
+
+		/* Figure out if extra threads are required. */
+		actual_id = f->t->d->aoeminor % ncpus;
+
+		if (!kts[actual_id].active) {
+			BUG_ON(id != 0);
+			mutex_lock(&ktio_spawn_lock);
+			if (!kts[actual_id].active
+				&& aoe_ktstart(&kts[actual_id]) == 0)
+				kts[actual_id].active = 1;
+			mutex_unlock(&ktio_spawn_lock);
+		}
+		spin_lock_irq(&iocq[id].lock);
 	}
 }
 
@@ -1311,7 +1334,7 @@ kthread(void *vp)
 	complete(&k->rendez);	/* tell spawner we're running */
 	do {
 		spin_lock_irq(k->lock);
-		more = k->fn();
+		more = k->fn(k->id);
 		if (!more) {
 			add_wait_queue(k->waitq, &wait);
 			__set_current_state(TASK_INTERRUPTIBLE);
@@ -1353,13 +1376,24 @@ aoe_ktstart(struct ktstate *k)
 static void
 ktcomplete(struct frame *f, struct sk_buff *skb)
 {
+	int id;
 	ulong flags;
 
 	f->r_skb = skb;
-	spin_lock_irqsave(&iocq.lock, flags);
-	list_add_tail(&f->head, &iocq.head);
-	spin_unlock_irqrestore(&iocq.lock, flags);
-	wake_up(&ktiowq);
+	id = f->t->d->aoeminor % ncpus;
+	spin_lock_irqsave(&iocq[id].lock, flags);
+	if (!kts[id].active) {
+		spin_unlock_irqrestore(&iocq[id].lock, flags);
+		/* The thread with id has not been spawned yet,
+		 * so delegate the work to the main thread and
+		 * try spawning a new thread.
+		 */
+		id = 0;
+		spin_lock_irqsave(&iocq[id].lock, flags);
+	}
+	list_add_tail(&f->head, &iocq[id].head);
+	spin_unlock_irqrestore(&iocq[id].lock, flags);
+	wake_up(&ktiowq[id]);
 }
 
 struct sk_buff *
@@ -1706,6 +1740,17 @@ aoe_failbuf(struct aoedev *d, struct buf *buf)
 void
 aoe_flush_iocq(void)
 {
+	int i;
+
+	for (i = 0; i < ncpus; i++) {
+		if (kts[i].active)
+			aoe_flush_iocq_by_index(i);
+	}
+}
+
+void
+aoe_flush_iocq_by_index(int id)
+{
 	struct frame *f;
 	struct aoedev *d;
 	LIST_HEAD(flist);
@@ -1713,9 +1758,9 @@ aoe_flush_iocq(void)
 	struct sk_buff *skb;
 	ulong flags;
 
-	spin_lock_irqsave(&iocq.lock, flags);
-	list_splice_init(&iocq.head, &flist);
-	spin_unlock_irqrestore(&iocq.lock, flags);
+	spin_lock_irqsave(&iocq[id].lock, flags);
+	list_splice_init(&iocq[id].head, &flist);
+	spin_unlock_irqrestore(&iocq[id].lock, flags);
 	while (!list_empty(&flist)) {
 		pos = flist.next;
 		list_del(pos);
@@ -1738,6 +1783,8 @@ int __init
 aoecmd_init(void)
 {
 	void *p;
+	int i;
+	int ret;
 
 	/* get_zeroed_page returns page with ref count 1 */
 	p = (void *) get_zeroed_page(GFP_KERNEL | __GFP_REPEAT);
@@ -1745,22 +1792,76 @@ aoecmd_init(void)
 		return -ENOMEM;
 	empty_page = virt_to_page(p);
 
-	INIT_LIST_HEAD(&iocq.head);
-	spin_lock_init(&iocq.lock);
-	init_waitqueue_head(&ktiowq);
-	kts.name = "aoe_ktio";
-	kts.fn = ktio;
-	kts.waitq = &ktiowq;
-	kts.lock = &iocq.lock;
-	return aoe_ktstart(&kts);
+
+	/* The number of ktio threads to be spawned are governed by the
+	* number of online cpus available in the system.
+	*/
+	ncpus = num_online_cpus();
+
+	iocq = kcalloc(ncpus, sizeof(struct iocq_ktio), GFP_KERNEL);
+	if (!iocq)
+		return -ENOMEM;
+
+	kts = kcalloc(ncpus, sizeof(struct ktstate), GFP_KERNEL);
+	if (!kts) {
+		ret = -ENOMEM;
+		goto kts_fail;
+	}
+
+	ktiowq = kcalloc(ncpus, sizeof(wait_queue_head_t), GFP_KERNEL);
+	if (!ktiowq) {
+		ret = -ENOMEM;
+		goto ktiowq_fail;
+	}
+
+	mutex_init(&ktio_spawn_lock);
+
+	for (i = 0; i < ncpus; i++) {
+		INIT_LIST_HEAD(&iocq[i].head);
+		spin_lock_init(&iocq[i].lock);
+		init_waitqueue_head(&ktiowq[i]);
+		snprintf(kts[i].name, sizeof(kts[i].name), "aoe_ktio%d", i);
+		kts[i].fn = ktio;
+		kts[i].waitq = &ktiowq[i];
+		kts[i].lock = &iocq[i].lock;
+		kts[i].id = i;
+		kts[i].active = 0;
+	}
+	kts[0].active = 1;
+	if (aoe_ktstart(&kts[0])) {
+		ret = -ENOMEM;
+		goto ktstart_fail;
+	}
+	return 0;
+
+ktstart_fail:
+	kfree(ktiowq);
+ktiowq_fail:
+	kfree(kts);
+kts_fail:
+	kfree(iocq);
+
+	return ret;
 }
 
 void
 aoecmd_exit(void)
 {
-	aoe_ktstop(&kts);
+	int i;
+
+	for (i = 0; i < ncpus; i++)
+		if (kts[i].active)
+			aoe_ktstop(&kts[i]);
+
 	aoe_flush_iocq();
 
+	/* Free up the iocq and thread speicific configuration
+	* allocated during startup.
+	*/
+	kfree(iocq);
+	kfree(kts);
+	kfree(ktiowq);
+
 	free_page((unsigned long) page_address(empty_page));
 	empty_page = NULL;
 }
diff --git a/drivers/block/aoe/aoedev.c b/drivers/block/aoe/aoedev.c
index 98f2965..92201b6 100644
--- a/drivers/block/aoe/aoedev.c
+++ b/drivers/block/aoe/aoedev.c
@@ -518,7 +518,6 @@ void
 aoedev_exit(void)
 {
 	flush_scheduled_work();
-	aoe_flush_iocq();
 	flush(NULL, 0, EXITING);
 }
 
diff --git a/drivers/block/aoe/aoenet.c b/drivers/block/aoe/aoenet.c
index 71d3ea8..4af5f06 100644
--- a/drivers/block/aoe/aoenet.c
+++ b/drivers/block/aoe/aoenet.c
@@ -52,7 +52,7 @@ static struct sk_buff_head skbtxq;
 
 /* enters with txlock held */
 static int
-tx(void) __must_hold(&txlock)
+tx(int id) __must_hold(&txlock)
 {
 	struct sk_buff *skb;
 	struct net_device *ifp;
@@ -205,7 +205,8 @@ aoenet_init(void)
 	kts.lock = &txlock;
 	kts.fn = tx;
 	kts.waitq = &txwq;
-	kts.name = "aoe_tx";
+	kts.id = 0;
+	snprintf(kts.name, sizeof(kts.name), "aoe_tx%d", kts.id);
 	if (aoe_ktstart(&kts))
 		return -EAGAIN;
 	dev_add_pack(&aoe_pt);
-- 
1.7.1

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists