Skip to contentSkip to conversation

bonding: simplify and use RCU protection for 3ad xmit path

PATCH Superseded Ding Tianhong 4 participants 5 messages Sep 4, 2013 – Sep 5, 2013

Subsystems (3)
Networking Network drivers Bonding driver

v2 of 5 Superseded Latest is v5 →

lore
Share & tools
lore

Ding Tianhong

author
#0 lore raw

PATCH v2 1/6 bonding: simplify and use RCU protection for 3ad xmit path

The commit 278b20837511776dc9d5f6ee1c7fabd5479838bb
(bonding: initial RCU conversion) has convert the roundrobin, active-backup, broadcast and xor xmit path to rcu protection, the performance will be better for these mode, so this time, convert xmit path for 3ad mode.

Suggested-by Nikolay Aleksandrov <[email protected]>
Signed-off-by Ding Tianhong <[email protected]>
Signed-off-by Wang Yufen <[email protected]>
Cc Nikolay Aleksandrov <[email protected]>
Cc Veaceslav Falico <[email protected]>
---
 drivers/net/bonding/bond_3ad.c | 32 ++++++++++++++------------------
 drivers/net/bonding/bonding.h  | 22 ++++++++++++++++++++++
 2 files changed, 36 insertions(+), 18 deletions(-)
drivers/net/bonding/bond_3ad.c +14 -18
index 9010265..7a3860f 100644
@@ -143,7 +143,7 @@ static inline struct bonding *__get_bond_by_port(struct port *port)
  */
 static inline struct port *__get_first_port(struct bonding *bond)
 {
-	struct slave *first_slave = bond_first_slave(bond);
+	struct slave *first_slave = bond_first_slave_rcu(bond);
 
 	return first_slave ? &(SLAVE_AD_INFO(first_slave).port) : NULL;
 }

@@ -163,7 +163,7 @@ static inline struct port *__get_next_port(struct port *port)
 	// If there's no bond for this port, or this is the last slave
 	if (bond == NULL)
 		return NULL;
-	slave_next = bond_next_slave(bond, slave);
+	slave_next = bond_next_slave_rcu(bond, slave);
 	if (!slave_next || bond_is_first_slave(bond, slave_next))
 		return NULL;
 

@@ -2417,16 +2417,14 @@ int bond_3ad_get_active_agg_info(struct bonding *bond, struct ad_info *ad_info)
 
 int bond_3ad_xmit_xor(struct sk_buff *skb, struct net_device *dev)
 {
-	struct slave *slave, *start_at;
 	struct bonding *bond = netdev_priv(dev);
+	struct slave *slave;
 	int slave_agg_no;
 	int slaves_in_agg;
 	int agg_id;
-	int i;
 	struct ad_info ad_info;
 	int res = 1;
 
-	read_lock(&bond->lock);
 	if (__bond_3ad_get_active_agg_info(bond, &ad_info)) {
 		pr_debug("%s: Error: __bond_3ad_get_active_agg_info failed\n",
 			 dev->name);

@@ -2444,13 +2442,17 @@ int bond_3ad_xmit_xor(struct sk_buff *skb, struct net_device *dev)
 
 	slave_agg_no = bond->xmit_hash_policy(skb, slaves_in_agg);
 
-	bond_for_each_slave(bond, slave) {
+	bond_for_each_slave_rcu(bond, slave) {
 		struct aggregator *agg = SLAVE_AD_INFO(slave).port.aggregator;
 
 		if (agg && (agg->aggregator_identifier == agg_id)) {
-			slave_agg_no--;
-			if (slave_agg_no < 0)
-				break;
+			if (--slave_agg_no < 0) {
+				if (SLAVE_IS_OK(slave)) {
+					res = bond_dev_queue_xmit(bond,
+						skb, slave->dev);
+					goto out;
+				}
+			}
 		}
 	}
 

@@ -2460,23 +2462,17 @@ int bond_3ad_xmit_xor(struct sk_buff *skb, struct net_device *dev)
 		goto out;
 	}
 
-	start_at = slave;
-
-	bond_for_each_slave_from(bond, slave, i, start_at) {
-		int slave_agg_id = 0;
+	bond_for_each_slave_rcu(bond, slave) {
 		struct aggregator *agg = SLAVE_AD_INFO(slave).port.aggregator;
 
-		if (agg)
-			slave_agg_id = agg->aggregator_identifier;
-
-		if (SLAVE_IS_OK(slave) && agg && (slave_agg_id == agg_id)) {
+		if (SLAVE_IS_OK(slave) && agg &&
+			agg->aggregator_identifier == agg_id) {
 			res = bond_dev_queue_xmit(bond, skb, slave->dev);
 			break;
 		}
 	}
 
 out:
-	read_unlock(&bond->lock);
 	if (res) {
 		/* no suitable interface, frame not sent */
 		kfree_skb(skb);
drivers/net/bonding/bonding.h +22 -0
index 4bf52d5..9898493 100644
@@ -74,6 +74,9 @@
 /* slave list primitives */
 #define bond_to_slave(ptr) list_entry(ptr, struct slave, list)
 
+/* slave list primitives, Caller must hold rcu_read_lock */
+#define bond_to_slave_rcu(ptr) list_entry_rcu(ptr, struct slave, list)
+
 /* IMPORTANT: bond_first/last_slave can return NULL in case of an empty list */
 #define bond_first_slave(bond) \
 	list_first_entry_or_null(&(bond)->slave_list, struct slave, list)

@@ -81,6 +84,16 @@
 	(list_empty(&(bond)->slave_list) ? NULL : \
 					   bond_to_slave((bond)->slave_list.prev))
 
+/**
+ * IMPORTANT: bond_first/last_slave_rcu can return NULL in case of an empty list
+ * Caller must hold rcu_read_lock
+ */
+#define bond_first_slave_rcu(bond) \
+	list_first_or_null_rcu(&(bond)->slave_list, struct slave, list)
+#define bond_last_slave_rcu(bond) \
+	(list_empty(&(bond)->slave_list) ? NULL : \
+						bond_to_slave_rcu((bond)->slave_list.prev))
+
 #define bond_is_first_slave(bond, pos) ((pos)->list.prev == &(bond)->slave_list)
 #define bond_is_last_slave(bond, pos) ((pos)->list.next == &(bond)->slave_list)
 

@@ -93,6 +106,15 @@
 	(bond_is_first_slave(bond, pos) ? bond_last_slave(bond) : \
 					  bond_to_slave((pos)->list.prev))
 
+/* Since bond_first/last_slave_rcu can return NULL, these can return NULL too */
+#define bond_next_slave_rcu(bond, pos) \
+	(bond_is_last_slave(bond, pos) ? bond_first_slave_rcu(bond) : \
+					 bond_to_slave_rcu((pos)->list.next))
+
+#define bond_prev_slave_rcu(bond, pos) \
+	(bond_is_first_slave(bond, pos) ? bond_last_slave_rcu(bond) : \
+					  bond_to_slave_rcu((pos)->list.prev))
+
 /**
  * bond_for_each_slave_from - iterate the slaves list from a starting point
  * @bond:	the bond holding this list.
Signature · 1 line
1.8.2.1
#1 lore raw

On Wed, Sep 04, 2013 at 05:43:45PM +0800, Ding Tianhong wrote:
...snip...

drivers/net/bonding/bonding.h line 95

Quoted 9 lines
drivers/net/bonding/bonding.h +9 -0
+/**
+ * IMPORTANT: bond_first/last_slave_rcu can return NULL in case of an empty list
+ * Caller must hold rcu_read_lock
+ */
+#define bond_first_slave_rcu(bond) \
+	list_first_or_null_rcu(&(bond)->slave_list, struct slave, list)
+#define bond_last_slave_rcu(bond) \
+	(list_empty(&(bond)->slave_list) ? NULL : \
+						bond_to_slave_rcu((bond)->slave_list.prev))

Here, bond_last_slave_rcu() is racy. The list can be non-empty when list_empty() is verified, however afterwards it might become empty, when you call bond_to_slave_rcu(), and thus you'll get bond_to_slave(bond->slave_list) in the result, which is not a slave.

Take a look at list_first_or_null_rcu() for a reference. The main idea is that it first gets the ->next pointer, with RCU protection, and then verifies if it's the list head or not, and if not - it gets the container already. This way the ->next pointer won't get away.

These kind of bugs are really rare, but are *EXTREMELY* hard to debug.

Quoted 7 lines

+
#define bond_is_first_slave(bond, pos) ((pos)->list.prev == &(bond)->slave_list) #define bond_is_last_slave(bond, pos) ((pos)->list.next == &(bond)->slave_list)

drivers/net/bonding/bonding.h +0 -0
@@ -93,6 +106,15 @@
 	(bond_is_first_slave(bond, pos) ? bond_last_slave(bond) : \
 					  bond_to_slave((pos)->list.prev))

drivers/net/bonding/bonding.h line 117

Quoted 10 lines
drivers/net/bonding/bonding.h +9 -0
 
+/* Since bond_first/last_slave_rcu can return NULL, these can return NULL too */
+#define bond_next_slave_rcu(bond, pos) \
+	(bond_is_last_slave(bond, pos) ? bond_first_slave_rcu(bond) : \
+					 bond_to_slave_rcu((pos)->list.next))
+
+#define bond_prev_slave_rcu(bond, pos) \
+	(bond_is_first_slave(bond, pos) ? bond_last_slave_rcu(bond) : \
+					  bond_to_slave_rcu((pos)->list.prev))
+

These two are also racy. bond_is_last/first_slave() is not rcu-ified, and thus you can't rely on it without proper locking. Same ideas apply as per bond_first_slave_rcu().

#2 lore raw
Quoted 18 lines · Veaceslav Falico

On Wed, Sep 04, 2013 at 05:43:45PM +0800, Ding Tianhong wrote:
...snip...

drivers/net/bonding/bonding.h line 95

Quoted 10 lines

+/**
+ * IMPORTANT: bond_first/last_slave_rcu can return NULL in case of an empty list
+ * Caller must hold rcu_read_lock
+ */

drivers/net/bonding/bonding.h +5 -0
+#define bond_first_slave_rcu(bond) \
+ list_first_or_null_rcu(&(bond)->slave_list, struct slave, list)
+#define bond_last_slave_rcu(bond) \
+ (list_empty(&(bond)->slave_list) ? NULL : \
+ bond_to_slave_rcu((bond)->slave_list.prev))

Here, bond_last_slave_rcu() is racy. The list can be non-empty when list_empty() is verified, however afterwards it might become empty, when you call bond_to_slave_rcu(), and thus you'll get bond_to_slave(bond->slave_list) in the result, which is not a slave.

Quoted 6 lines

Take a look at list_first_or_null_rcu() for a reference. The main idea is that it first gets the ->next pointer, with RCU protection, and then verifies if it's the list head or not, and if not - it gets the container already. This way the ->next pointer won't get away.

These kind of bugs are really rare, but are *EXTREMELY* hard to debug.

Thanks for your response and opinions, but I think your miss something, the slave_list will not changed in the rcu_read_lock, so ,the bugs will not happen.

Quoted 9 lines
Quoted 9 lines

+
#define bond_is_first_slave(bond, pos) ((pos)->list.prev == &(bond)->slave_list) #define bond_is_last_slave(bond, pos) ((pos)->list.next == &(bond)->slave_list)

drivers/net/bonding/bonding.h +0 -0
@@ -93,6 +106,15 @@
 (bond_is_first_slave(bond, pos) ? bond_last_slave(bond) : \
 bond_to_slave((pos)->list.prev))
Quoted 15 lines

drivers/net/bonding/bonding.h line 117

Quoted 10 lines
(quoted diff) +1 -0
+/* Since bond_first/last_slave_rcu can return NULL, these can return

NULL too */

drivers/net/bonding/bonding.h +8 -0
+#define bond_next_slave_rcu(bond, pos) \
+ (bond_is_last_slave(bond, pos) ? bond_first_slave_rcu(bond) : \
+ bond_to_slave_rcu((pos)->list.next))
+
+#define bond_prev_slave_rcu(bond, pos) \
+ (bond_is_first_slave(bond, pos) ? bond_last_slave_rcu(bond) : \
+ bond_to_slave_rcu((pos)->list.prev))
+

These two are also racy. bond_is_last/first_slave() is not rcu-ified, and thus you can't rely on it without proper locking. Same ideas apply as per bond_first_slave_rcu().

refer to the above answer.

Quoted 4 lines

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

maintainer
#3 lore raw
In reply to Veaceslav Falico

From: Veaceslav Falico <[email protected]>
Date: Wed, 4 Sep 2013 12:18:24 +0200

Quoted 19 lines · Ding Tianhong

...snip...

drivers/net/bonding/bonding.h line 95

Quoted 10 lines

+/**
+ * IMPORTANT: bond_first/last_slave_rcu can return NULL in case of an empty list
+ * Caller must hold rcu_read_lock
+ */

drivers/net/bonding/bonding.h +5 -0
+#define bond_first_slave_rcu(bond) \
+	list_first_or_null_rcu(&(bond)->slave_list, struct slave, list)
+#define bond_last_slave_rcu(bond) \
+	(list_empty(&(bond)->slave_list) ? NULL : \
+ bond_to_slave_rcu((bond)->slave_list.prev))

Here, bond_last_slave_rcu() is racy. The list can be non-empty when list_empty() is verified, however afterwards it might become empty, when
you call bond_to_slave_rcu(), and thus you'll get bond_to_slave(bond->slave_list) in the result, which is not a slave.

Quoted 8 lines

Take a look at list_first_or_null_rcu() for a reference. The main idea is
that it first gets the ->next pointer, with RCU protection, and then verifies if it's the list head or not, and if not - it gets the container
already. This way the ->next pointer won't get away.

These kind of bugs are really rare, but are *EXTREMELY* hard to debug.

I agree with this analysis.

Ding, "rcu_read_lock()" doesn't "lock" anything. It's just a memory barrier.

All the list can still change on you asynchronously to your accesses.

That's why list_first_or_null_rcu() is so carefully arranged. Therefore, you must make similar accomodations.

author
#4 lore raw
Quoted 33 lines · David Miller

From: Veaceslav Falico <[email protected]>
Date: Wed, 4 Sep 2013 12:18:24 +0200

Quoted 19 lines · Ding Tianhong

...snip...

drivers/net/bonding/bonding.h line 95

Quoted 10 lines

+/**
+ * IMPORTANT: bond_first/last_slave_rcu can return NULL in case of an empty list
+ * Caller must hold rcu_read_lock
+ */

drivers/net/bonding/bonding.h +5 -0
+#define bond_first_slave_rcu(bond) \
+	list_first_or_null_rcu(&(bond)->slave_list, struct slave, list)
+#define bond_last_slave_rcu(bond) \
+	(list_empty(&(bond)->slave_list) ? NULL : \
+ bond_to_slave_rcu((bond)->slave_list.prev))

Here, bond_last_slave_rcu() is racy. The list can be non-empty when list_empty() is verified, however afterwards it might become empty, when
you call bond_to_slave_rcu(), and thus you'll get bond_to_slave(bond->slave_list) in the result, which is not a slave.

Quoted 8 lines

Take a look at list_first_or_null_rcu() for a reference. The main idea is
that it first gets the ->next pointer, with RCU protection, and then verifies if it's the list head or not, and if not - it gets the container
already. This way the ->next pointer won't get away.

These kind of bugs are really rare, but are *EXTREMELY* hard to debug.

I agree with this analysis.

Quoted 10 lines

Ding, "rcu_read_lock()" doesn't "lock" anything. It's just a memory barrier.

All the list can still change on you asynchronously to your accesses.

That's why list_first_or_null_rcu() is so carefully arranged. Therefore, you must make similar accomodations.

yes, after a long time thinking, I found the problem and know how to do next, repair and resend it later.

Quoted 2 lines

.