Skip to contentSkip to conversation

netdev: get rid of read_lock(&dev_base_lock) usages

PATCH Applied Stephen Hemminger 8 participants 59 messages Nov 10, 2009 – Nov 18, 2009

Subsystems (10)
lore
Share & tools
lore
#0 lore raw

cover netdev: get rid of read_lock(&dev_base_lock) usages

The goal is to eliminate dev_base_lock completely, and just use RCU and rtnl_mutex for network devices. This series gets rid of the many of the users of dev_base_lock just for reading.

#1 lore raw

PATCH 1/10 netdev: add netdev_continue_rcu

This adds an RCU macro for continuing search, useful for some network devices like vlan.

include/linux/netdevice.h +2 -0
@@ -1079,6 +1079,8 @@ extern rwlock_t				dev_base_lock;		/* De
 		list_for_each_entry_safe(d, n, &(net)->dev_base_head, dev_list)
 #define for_each_netdev_continue(net, d)		\
 		list_for_each_entry_continue(d, &(net)->dev_base_head, dev_list)
+#define for_each_netdev_continue_rcu(net, d)		\
+	list_for_each_entry_continue_rcu(d, &(net)->dev_base_head, dev_list)
 #define net_device_entry(lh)	list_entry(lh, struct net_device, dev_list)
 
 static inline struct net_device *next_net_device(struct net_device *dev)
include/linux/rculist.h +14 -0
@@ -262,6 +262,20 @@ static inline void list_splice_init_rcu(
 		(pos) = rcu_dereference((pos)->next))
 
 /**
+ * list_for_each_entry_continue_rcu - continue iteration over list of given type
+ * @pos:	the type * to use as a loop cursor.
+ * @head:	the head for your list.
+ * @member:	the name of the list_struct within the struct.
+ *
+ * Continue to iterate over list of given type, continuing after
+ * the current position.
+ */
+#define list_for_each_entry_continue_rcu(pos, head, member) 		\
+	for (pos = list_entry_rcu(pos->member.next, typeof(*pos), member); \
+	     prefetch(pos->member.next), &pos->member != (head);	\
+	     pos = list_entry_rcu(pos->member.next, typeof(*pos), member))
+
+/**
  * hlist_del_rcu - deletes entry from hash list without re-initialization
  * @n: the element to delete from the hash list.
  *
#4 lore raw

Re: [PATCH 01/10] netdev: add netdev_continue_rcu

In reply to Stephen Hemminger
Quoted 2 lines · Stephen Hemminger

This adds an RCU macro for continuing search, useful for some network devices like vlan.

Looks good!!!

Of course, you need to either have a single RCU read-side critical section cover all the chained list_for_each_entry_continue_rcu() invocations, or you need to do something (e.g., reference count) to make sure that the element in question doesn't disappear in the meantime, right?

Thanx, Paul

Quoted 39 lines
include/linux/netdevice.h +2 -0
@@ -1079,6 +1079,8 @@ extern rwlock_t				dev_base_lock;		/* De
 		list_for_each_entry_safe(d, n, &(net)->dev_base_head, dev_list)
 #define for_each_netdev_continue(net, d)		\
 		list_for_each_entry_continue(d, &(net)->dev_base_head, dev_list)
+#define for_each_netdev_continue_rcu(net, d)		\
+	list_for_each_entry_continue_rcu(d, &(net)->dev_base_head, dev_list)
 #define net_device_entry(lh)	list_entry(lh, struct net_device, dev_list)
 
 static inline struct net_device *next_net_device(struct net_device *dev)
include/linux/rculist.h +14 -0
@@ -262,6 +262,20 @@ static inline void list_splice_init_rcu(
 		(pos) = rcu_dereference((pos)->next))
 
 /**
+ * list_for_each_entry_continue_rcu - continue iteration over list of given type
+ * @pos:	the type * to use as a loop cursor.
+ * @head:	the head for your list.
+ * @member:	the name of the list_struct within the struct.
+ *
+ * Continue to iterate over list of given type, continuing after
+ * the current position.
+ */
+#define list_for_each_entry_continue_rcu(pos, head, member) 		\
+	for (pos = list_entry_rcu(pos->member.next, typeof(*pos), member); \
+	     prefetch(pos->member.next), &pos->member != (head);	\
+	     pos = list_entry_rcu(pos->member.next, typeof(*pos), member))
+
+/**
  * hlist_del_rcu - deletes entry from hash list without re-initialization
  * @n: the element to delete from the hash list.
  *
Signature · 1 line
-- 
#5 lore raw

PATCH 2/10 vlan: eliminate use of dev_base_lock

In reply to Stephen Hemminger

Do not need to use read_lock(&dev_base_lock), use RCU instead.

net/8021q/vlanproc.c +6 -7
@@ -201,18 +201,17 @@ int vlan_proc_rem_dev(struct net_device 
 
 /* start read of /proc/net/vlan/config */
 static void *vlan_seq_start(struct seq_file *seq, loff_t *pos)
-	__acquires(dev_base_lock)
+	__acquires(rcu)
 {
 	struct net_device *dev;
 	struct net *net = seq_file_net(seq);
 	loff_t i = 1;
 
-	read_lock(&dev_base_lock);
-
+	rcu_read_lock();
 	if (*pos == 0)
 		return SEQ_START_TOKEN;
 
-	for_each_netdev(net, dev) {
+	for_each_netdev_rcu(net, dev) {
 		if (!is_vlan_dev(dev))
 			continue;
 
@@ -234,7 +233,7 @@ static void *vlan_seq_next(struct seq_fi
 	if (v == SEQ_START_TOKEN)
 		dev = net_device_entry(&net->dev_base_head);
 
-	for_each_netdev_continue(net, dev) {
+	for_each_netdev_continue_rcu(net, dev) {
 		if (!is_vlan_dev(dev))
 			continue;
 
@@ -245,9 +244,9 @@ static void *vlan_seq_next(struct seq_fi
 }
 
 static void vlan_seq_stop(struct seq_file *seq, void *v)
-	__releases(dev_base_lock)
+	__releases(rcu)
 {
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 }
 
 static int vlan_seq_show(struct seq_file *seq, void *v)
#8 lore raw

PATCH 3/10 net: use rcu for network scheduler API

In reply to Stephen Hemminger

Use RCU to walk list of network devices in qdisc dump. This could be optimized for large number of devices.

net/sched/sch_api.c +4 -3
@@ -1279,9 +1279,10 @@ static int tc_dump_qdisc(struct sk_buff 
 
 	s_idx = cb->args[0];
 	s_q_idx = q_idx = cb->args[1];
-	read_lock(&dev_base_lock);
+
+	rcu_read_lock();
 	idx = 0;
-	for_each_netdev(&init_net, dev) {
+	for_each_netdev_rcu(&init_net, dev) {
 		struct netdev_queue *dev_queue;
 
 		if (idx < s_idx)
@@ -1302,7 +1303,7 @@ cont:
 	}
 
 done:
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 
 	cb->args[0] = idx;
 	cb->args[1] = q_idx;
#11 lore raw

PATCH 4/10 AOE: use rcu to find network device

In reply to Stephen Hemminger

This gets rid of another use of read_lock(&dev_base_lock) by using RCU. Also, it only increments the reference count of the device actually used rather than holding and releasing every device

Compile tested only.

drivers/block/aoe/aoecmd.c +7 -9
@@ -296,17 +296,18 @@ aoecmd_cfg_pkts(ushort aoemajor, unsigne
 	struct sk_buff *skb;
 	struct net_device *ifp;
 
-	read_lock(&dev_base_lock);
-	for_each_netdev(&init_net, ifp) {
-		dev_hold(ifp);
+	rcu_read_lock();
+	for_each_netdev_rcu(&init_net, ifp) {
 		if (!is_aoe_netif(ifp))
-			goto cont;
+			continue;
 
 		skb = new_skb(sizeof *h + sizeof *ch);
 		if (skb == NULL) {
 			printk(KERN_INFO "aoe: skb alloc failure\n");
-			goto cont;
+			continue;
 		}
+
+		dev_hold(ifp);
 		skb_put(skb, sizeof *h + sizeof *ch);
 		skb->dev = ifp;
 		__skb_queue_tail(queue, skb);
@@ -320,11 +321,8 @@ aoecmd_cfg_pkts(ushort aoemajor, unsigne
 		h->major = cpu_to_be16(aoemajor);
 		h->minor = aoeminor;
 		h->cmd = AOECMD_CFG;
-
-cont:
-		dev_put(ifp);
 	}
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 }
 
 static void
#13 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

In reply to Stephen Hemminger
Quoted 5 lines · On Tue Nov 10 13:07:37 EST 2009, [email protected] wrote:

This gets rid of another use of read_lock(&dev_base_lock) by using RCU. Also, it only increments the reference count of the device actually used rather than holding and releasing every device

Compile tested only.

This function runs once a minute when the aoe driver is loaded, if you'd like to test it a bit more.

It looks like there's no dev_put corresponding to the dev_hold after the changes.

Signature · 1 line
  Ed
#14 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

Quoted 5 lines · On Tue, 10 Nov 2009 15:01:49 -0500 Ed Cashin <[email protected]> wrote:
Quoted 4 lines · On Tue Nov 10 13:07:37 EST 2009, [email protected] wrote:

This gets rid of another use of read_lock(&dev_base_lock) by using RCU. Also, it only increments the reference count of the device actually used rather than holding and releasing every device

Quoted 8 lines
Quoted 1 line

Compile tested only.

This function runs once a minute when the aoe driver is loaded, if you'd like to test it a bit more.

It looks like there's no dev_put corresponding to the dev_hold after the changes.

Hmm, looks like AOE actually is not ref counting the network device. So my patch is incorrect.

As it stands (before my patch), it is UNSAFE. It can decide to queue packets to a device that is removed out from underneath it causing reference to freed memory.

Moving the rcu_read_lock up to aoecmd_cfg() would solve that but the whole driver appears to be unsafe about device refcounting and handling device removal properly.

It needs to:

1. Get a device ref count when it remembers a device: (ie addif)
2. Install a notifier that looks for device removal events
3. In notifier, remove interface, including flushing all pending skb's for that device.

This obviously is beyond the scope of the RCU stuff.

#15 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

Quoted 28 lines · On Tue, 10 Nov 2009 15:06:17 -0800 Stephen Hemminger <[email protected]>…
Quoted 5 lines · On Tue, 10 Nov 2009 15:01:49 -0500 Ed Cashin <[email protected]> wrote:
Quoted 4 lines · On Tue Nov 10 13:07:37 EST 2009, [email protected] wrote:

This gets rid of another use of read_lock(&dev_base_lock) by using RCU. Also, it only increments the reference count of the device actually used rather than holding and releasing every device

Quoted 8 lines
Quoted 1 line

Compile tested only.

This function runs once a minute when the aoe driver is loaded, if you'd like to test it a bit more.

It looks like there's no dev_put corresponding to the dev_hold after the changes.

Hmm, looks like AOE actually is not ref counting the network device. So my patch is incorrect.

As it stands (before my patch), it is UNSAFE. It can decide to queue packets to a device that is removed out from underneath it causing reference to freed memory.

Moving the rcu_read_lock up to aoecmd_cfg() would solve that but the whole driver appears to be unsafe about device refcounting and handling device removal properly.

Quoted 8 lines

It needs to:

1. Get a device ref count when it remembers a device: (ie addif)
2. Install a notifier that looks for device removal events
3. In notifier, remove interface, including flushing all pending skb's for that device.

This obviously is beyond the scope of the RCU stuff.

Here is a patch to get you going, it does compile but it probably won't work because the code doesn't handle the case of the last device going away from a target. This is yet another pre-existing bug, since if a timeout happens: ejectif() is called to remove a device, resend() will BUG in ifrotate() if all devices are gone.

---
 drivers/block/aoe/aoe.h     |    2 ++
 drivers/block/aoe/aoecmd.c  |   19 +++++++++++++++++++
 drivers/block/aoe/aoedev.c  |   14 ++++++++++++++
 drivers/block/aoe/aoemain.c |   24 ++++++++++++++++++++++++
 4 files changed, 59 insertions(+)
drivers/block/aoe/aoecmd.c +19 -0
@@ -413,6 +413,8 @@ addif(struct aoetgt *t, struct net_devic
 	p = getif(t, NULL);
 	if (!p)
 		return NULL;
+
+	dev_hold(nd);
 	p->nd = nd;
 	p->maxbcnt = DEFAULTBCNT;
 	p->lost = 0;
@@ -424,12 +426,29 @@ static void
 ejectif(struct aoetgt *t, struct aoeif *ifp)
 {
 	struct aoeif *e;
+	struct net_device *nd;
 	ulong n;
 
 	e = t->ifs + NAOEIFS - 1;
+	nd = e->nd;
 	n = (e - ifp) * sizeof *ifp;
 	memmove(ifp, ifp+1, n);
 	e->nd = NULL;
+	dev_put(nd);
+}
+
+void aoecmd_flushnet(struct aoedev *d, struct net_device *nd)
+{
+	struct aoetgt **tt, **te;
+	tt = d->targets;
+	te = tt + NTARGETS;
+	for (; tt < te && *tt; tt++) {
+		struct aoetgt *t = *tt;
+		struct aoeif *ifp;
+
+		while ( (ifp = getif(t, nd)) )
+			ejectif(t, ifp);
+	}
 }
 
 static int
drivers/block/aoe/aoemain.c +24 -0
@@ -8,6 +8,8 @@
 #include <linux/blkdev.h>
 #include <linux/module.h>
 #include <linux/skbuff.h>
+#include <linux/notifier.h>
+#include <linux/netdevice.h>
 #include "aoe.h"
 
 MODULE_LICENSE("GPL");
@@ -54,11 +56,28 @@ discover_timer(ulong vp)
 	}
 }
 
+/* Callback on change of state of network device. */
+static int aoe_device_event(struct notifier_block *unused,
+			    unsigned long event, void *ptr)
+{
+	struct net_device *nd = ptr;
+
+	if (is_aoe_netif(nd) && event == NETDEV_UNREGISTER)
+		aoedev_ejectnet(nd);
+
+	return NOTIFY_DONE;
+}
+
+static struct notifier_block aoe_notifier = {
+	.notifier_call = aoe_device_event,
+};
+
 static void
 aoe_exit(void)
 {
 	discover_timer(TKILL);
 
+	unregister_netdevice_notifier(&aoe_notifier);
 	aoenet_exit();
 	unregister_blkdev(AOE_MAJOR, DEVICE_NAME);
 	aoechr_exit();
@@ -83,6 +102,9 @@ aoe_init(void)
 	ret = aoenet_init();
 	if (ret)
 		goto net_fail;
+	ret = register_netdevice_notifier(&aoe_notifier);
+	if (ret)
+		goto notifier_fail;
 	ret = register_blkdev(AOE_MAJOR, DEVICE_NAME);
 	if (ret < 0) {
 		printk(KERN_ERR "aoe: can't register major\n");
@@ -94,6 +116,8 @@ aoe_init(void)
 	return 0;
 
  blkreg_fail:
+	unregister_netdevice_notifier(&aoe_notifier);
+ notifier_fail:
 	aoenet_exit();
  net_fail:
 	aoeblk_exit();
drivers/block/aoe/aoe.h +2 -0
@@ -186,6 +186,7 @@ void aoecmd_ata_rsp(struct sk_buff *);
 void aoecmd_cfg_rsp(struct sk_buff *);
 void aoecmd_sleepwork(struct work_struct *);
 void aoecmd_cleanslate(struct aoedev *);
+void aoecmd_flushnet(struct aoedev *, struct net_device *);
 struct sk_buff *aoecmd_ata_id(struct aoedev *);
 
 int aoedev_init(void);
@@ -194,6 +195,7 @@ struct aoedev *aoedev_by_aoeaddr(int maj
 struct aoedev *aoedev_by_sysminor_m(ulong sysminor);
 void aoedev_downdev(struct aoedev *d);
 int aoedev_flush(const char __user *str, size_t size);
+void aoedev_ejectnet(struct net_device *);
 
 int aoenet_init(void);
 void aoenet_exit(void);
drivers/block/aoe/aoedev.c +14 -0
@@ -162,6 +162,20 @@ aoedev_flush(const char __user *str, siz
 	return 0;
 }
 
+void aoedev_ejectnet(struct net_device *nd)
+{
+	struct aoedev *d;
+	unsigned long flags;
+
+	spin_lock_irqsave(&devlist_lock, flags);
+	for (d = devlist; d; d = d->next) {
+		spin_lock(&d->lock);
+		aoecmd_flushnet(d, nd);
+		spin_unlock(&d->lock);
+	}
+	spin_unlock_irqrestore(&d->lock, flags);
+}
+
 /* I'm not really sure that this is a realistic problem, but if the
 network driver goes gonzo let's just leak memory after complaining. */
 static void
maintainer
#16 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

From: Stephen Hemminger <[email protected]>
Date: Tue, 10 Nov 2009 15:53:16 -0800

Quoted 38 lines
Quoted 35 lines · On Tue, 10 Nov 2009 15:06:17 -0800 Stephen Hemminger <[email protected]>…
Quoted 5 lines · On Tue, 10 Nov 2009 15:01:49 -0500 Ed Cashin <[email protected]> wrote:
Quoted 4 lines · On Tue Nov 10 13:07:37 EST 2009, [email protected] wrote:

This gets rid of another use of read_lock(&dev_base_lock) by using RCU. Also, it only increments the reference count of the device actually used rather than holding and releasing every device

Quoted 8 lines
Quoted 1 line

Compile tested only.

This function runs once a minute when the aoe driver is loaded, if you'd like to test it a bit more.

It looks like there's no dev_put corresponding to the dev_hold after the changes.

Hmm, looks like AOE actually is not ref counting the network device. So my patch is incorrect.

As it stands (before my patch), it is UNSAFE. It can decide to queue packets to a device that is removed out from underneath it causing reference to freed memory.

Moving the rcu_read_lock up to aoecmd_cfg() would solve that but the whole driver appears to be unsafe about device refcounting and handling device removal properly.

It needs to:

1. Get a device ref count when it remembers a device: (ie addif)
2. Install a notifier that looks for device removal events
3. In notifier, remove interface, including flushing all pending skb's for that device.

Quoted 7 lines
Quoted 1 line

This obviously is beyond the scope of the RCU stuff.

Here is a patch to get you going, it does compile but it probably won't work because the code doesn't handle the case of the last device going away from a target. This is yet another pre-existing bug, since if a timeout happens: ejectif() is called to remove a device, resend() will BUG in ifrotate() if all devices are gone.

I'm holding off on the RCU patch until these known refcount bugs are fixed

#17 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

In reply to Stephen Hemminger

Thanks again for providing this patch to help get things started. It's very helpful. I appreciate the way it reflects and fits into the rest of the driver, too.

In the patch, there's a loop around getif, ejectif inside the new aoecmd_flushnet function:

  void aoecmd_flushnet(struct aoedev *d, struct net_device *nd)
  {
  	struct aoetgt **tt, **te;
  	tt = d->targets;
  	te = tt + NTARGETS;
  	for (; tt < te && *tt; tt++) {
  		struct aoetgt *t = *tt;
  		struct aoeif *ifp;
  
  		while ( (ifp = getif(t, nd)) )
  			ejectif(t, ifp);
  	}
  }

... but an "if" seems appropriate, since duplicates are avoided when network devices are added in aoecmd_cfg_rsp:

  	ifp = getif(t, skb->dev);
  	if (!ifp) {
  		ifp = addif(t, skb->dev);

If there's some other reason for it to be "while" in aoecmd_flushnet that I'm missing, please let me know.

Regarding the new aoe_device_event function,

/* Callback on change of state of network device. */ static int aoe_device_event(struct notifier_block *unused,
unsigned long event, void *ptr)
{
struct net_device *nd = ptr;

  	if (is_aoe_netif(nd) && event == NETDEV_UNREGISTER)
  		aoedev_ejectnet(nd);
  
  	return NOTIFY_DONE;
  }

... the is_aoe_netif function really answers the question, "Are we allowed by the user to do AoE on this local network interface?" The user can modify the list of allowable interfaces at runtime via sysfs, as it's a module parameter. So I don't think we can use is_aoe_netif in the new aoe_device_event function. We should eject a net_device in this handler even if the user has decided not to use it for AoE. Please let me know if I'm missing the point.

You said that,

Quoted 6 lines

It needs to:

1. Get a device ref count when it remembers a device: (ie addif)
2. Install a notifier that looks for device removal events
3. In notifier, remove interface, including flushing all pending skb's for that device.

For 3, does the starter patch flush all pending skbs? Perhaps you could elaborate on what you had in mind?

I am trying to find the best way for the aoe driver to handle the situation where ther are no usable local network interfaces. It does seem to me that the user would benefit from a notice letting them know that they're trying to do AoE without any usable ethernet. I'm thinking that doing something like a printk_once would be appropriate.

(But probably not printk_once itself, because if they add a local interface and then it goes away later, they should be told again that something's wrong.)

Signature · 3 lines
  Ed Cashin <[email protected]>
  http://www.coraid.com/
  http://noserose.net/e/
#18 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

Quoted 21 lines · On Thu, 12 Nov 2009 09:33:16 -0500 Ed Cashin <[email protected]> wrote:

Thanks again for providing this patch to help get things started. It's very helpful. I appreciate the way it reflects and fits into the rest of the driver, too.

In the patch, there's a loop around getif, ejectif inside the new aoecmd_flushnet function:

  void aoecmd_flushnet(struct aoedev *d, struct net_device *nd)
  {
  	struct aoetgt **tt, **te;
  	tt = d->targets;
  	te = tt + NTARGETS;
  	for (; tt < te && *tt; tt++) {
  		struct aoetgt *t = *tt;
  		struct aoeif *ifp;
  
  		while ( (ifp = getif(t, nd)) )
  			ejectif(t, ifp);
  	}
  }
Quoted 6 lines

... but an "if" seems appropriate, since duplicates are avoided when network devices are added in aoecmd_cfg_rsp:

  	ifp = getif(t, skb->dev);
  	if (!ifp) {
  		ifp = addif(t, skb->dev);

Yeah if works, wasn't sure if you could have multiples.

Quoted 18 lines

If there's some other reason for it to be "while" in aoecmd_flushnet that I'm missing, please let me know.

Regarding the new aoe_device_event function,

/* Callback on change of state of network device. */ static int aoe_device_event(struct notifier_block *unused,
unsigned long event, void *ptr)
{
struct net_device *nd = ptr;

  	if (is_aoe_netif(nd) && event == NETDEV_UNREGISTER)
  		aoedev_ejectnet(nd);
  
  	return NOTIFY_DONE;
  }
Quoted 7 lines

... the is_aoe_netif function really answers the question, "Are we allowed by the user to do AoE on this local network interface?" The user can modify the list of allowable interfaces at runtime via sysfs, as it's a module parameter. So I don't think we can use is_aoe_netif in the new aoe_device_event function. We should eject a net_device in this handler even if the user has decided not to use it for AoE. Please let me know if I'm missing the point.

Ok, you know the code better, I just wanted to avoid doing unnecessary work.

Quoted 11 lines

You said that,

Quoted 6 lines

It needs to:

1. Get a device ref count when it remembers a device: (ie addif)
2. Install a notifier that looks for device removal events
3. In notifier, remove interface, including flushing all pending skb's for that device.

For 3, does the starter patch flush all pending skbs? Perhaps you could elaborate on what you had in mind?

I wasn't sure if you ended up queuing skb's, but it looks like the code queues requests not skb's. But you may need to disable preempt over code the work queue code that processes requests, to make sure that device doesn't disappear while doing stuff.

Quoted 6 lines

I am trying to find the best way for the aoe driver to handle the situation where ther are no usable local network interfaces. It does seem to me that the user would benefit from a notice letting them know that they're trying to do AoE without any usable ethernet. I'm thinking that doing something like a printk_once would be appropriate.

Since it is emulating a block device, why not propgate error back up the stack like a disk that's offline.

#19 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

Thanks for the responses.

On Thu Nov 12 12:11:12 EST 2009, [email protected] wrote:
...

Quoted 2 lines

Since it is emulating a block device, why not propgate error back up the stack like a disk that's offline.

The lack of local interfaces to use for AoE might be temporary. For example, the admin might be loading a new network driver, or a new link might go online through which the AoE target can be reached.

If AoE command packets are not sent or resent but instead are effectively dropped while there are no local interfaces through which to send, then the AoE commands will timeout through the normal mechanisms, at which time the error will be signaled to the block layer. It will be like normal unreliable ethernet transport.

In the common case, I think that's going to be the expected behavior, but a printk would probably still be helpful in case the admin doesn't realize why the AoE device is timing out.

Signature · 3 lines
  Ed Cashin <[email protected]>
  http://www.coraid.com/
  http://noserose.net/e/
#20 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

Quoted 17 lines · On Thu, 12 Nov 2009 13:07:35 -0500 Ed Cashin <[email protected]> wrote:

Thanks for the responses.

On Thu Nov 12 12:11:12 EST 2009, [email protected] wrote:
...

Quoted 2 lines

Since it is emulating a block device, why not propgate error back up the stack like a disk that's offline.

The lack of local interfaces to use for AoE might be temporary. For example, the admin might be loading a new network driver, or a new link might go online through which the AoE target can be reached.

If AoE command packets are not sent or resent but instead are effectively dropped while there are no local interfaces through which to send, then the AoE commands will timeout through the normal mechanisms, at which time the error will be signaled to the block layer. It will be like normal unreliable ethernet transport.

Quoted 4 lines

In the common case, I think that's going to be the expected behavior, but a printk would probably still be helpful in case the admin doesn't realize why the AoE device is timing out.

Okay, you might want to ratelimit the printk

#21 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

In reply to Stephen Hemminger

On Tue Nov 10 18:53:43 EST 2009, [email protected] wrote:

...

Quoted 2 lines
drivers/block/aoe/aoecmd.c +0 -0

...

Quoted 4 lines
drivers/block/aoe/aoecmd.c +0 -0
@@ -424,12 +426,29 @@ static void
 ejectif(struct aoetgt *t, struct aoeif *ifp)
 {
 	struct aoeif *e;

drivers/block/aoe/aoecmd.c static void · line 438

Quoted 10 lines
drivers/block/aoe/aoecmd.c +4 -0
+	struct net_device *nd;
 	ulong n;
 
 	e = t->ifs + NAOEIFS - 1;
+	nd = e->nd;
 	n = (e - ifp) * sizeof *ifp;
 	memmove(ifp, ifp+1, n);
 	e->nd = NULL;
+	dev_put(nd);
+}

I didn't notice this before, but e->nd will be the net_device pointer from the last ifp in the allocated array, whether it's NULL or not, and usually won't be ifp->nd, the one we need to give to dev_put. I think that's why I see an oops when testing the version of the patch that I sent last. The test is doing rmmod on the interface that's being used to create an ext2 fs on an AoE target.

I'll use ifp->nd instead of e->nd and test again tomorrow. I have to step away from the computer today.

Signature · 1 line
  Ed
#22 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

In reply to Stephen Hemminger

On Tue Nov 10 18:06:41 EST 2009, [email protected] wrote:
...

Quoted 6 lines

Hmm, looks like AOE actually is not ref counting the network device. So my patch is incorrect.

As it stands (before my patch), it is UNSAFE. It can decide to queue packets to a device that is removed out from underneath it causing reference to freed memory.

Yes, that's right. I will look at the patch you sent in the follow up. Thanks.

Signature · 1 line
  Ed
#23 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

I will be testing the patch below on Monday. It is based on Stephen Hemminger's starter patch (Thanks!). It adds handling for the case where all of the local network interfaces have become unusable. In that case, sent packets and retransmitted packets are effectively dropped, and the driver's handling of ethernet's unreliability takes care of things as usual. If a local interface becomes usable in time no I/O will fail.

After testing I'll make a proper changelog entry. In brief this patch allows the aoe driver to correctly take references on the net_device's that it uses, handle their removal via a callback, and handle their complete absence, too.

Signature · 1 line
  Ed
drivers/block/aoe/aoe.h +2 -0
index db195ab..315d36b 100644
@@ -186,6 +186,7 @@ void aoecmd_ata_rsp(struct sk_buff *);
 void aoecmd_cfg_rsp(struct sk_buff *);
 void aoecmd_sleepwork(struct work_struct *);
 void aoecmd_cleanslate(struct aoedev *);
+void aoecmd_flushnet(struct aoedev *, struct net_device *);
 struct sk_buff *aoecmd_ata_id(struct aoedev *);
 
 int aoedev_init(void);
@@ -194,6 +195,7 @@ struct aoedev *aoedev_by_aoeaddr(int maj, int min);
 struct aoedev *aoedev_by_sysminor_m(ulong sysminor);
 void aoedev_downdev(struct aoedev *d);
 int aoedev_flush(const char __user *str, size_t size);
+void aoedev_ejectnet(struct net_device *);
 
 int aoenet_init(void);
 void aoenet_exit(void);
drivers/block/aoe/aoecmd.c +27 -6
index 965ece2..060c8d6 100644
@@ -93,16 +93,16 @@ put_lba(struct aoe_atahdr *ah, sector_t lba)
 	ah->lba5 = lba >>= 8;
 }
 
-static void
+static struct net_device *
 ifrotate(struct aoetgt *t)
 {
 	t->ifp++;
 	if (t->ifp >= &t->ifs[NAOEIFS] || t->ifp->nd == NULL)
 		t->ifp = t->ifs;
-	if (t->ifp->nd == NULL) {
-		printk(KERN_INFO "aoe: no interface to rotate to\n");
-		BUG();
-	}
+	if (t->ifp->nd == NULL && net_ratelimit())
+		printk(KERN_WARNING
+			"aoe: no local interface to send through\n");
+	return t->ifp->nd;
 }
 
 static void
@@ -336,7 +336,8 @@ resend(struct aoedev *d, struct aoetgt *t, struct frame *f)
 	char buf[128];
 	u32 n;
 
-	ifrotate(t);
+	if (!ifrotate(t))
+		return;		/* no usable local network interfaces */
 	n = newtag(t);
 	skb = f->skb;
 	h = (struct aoe_hdr *) skb_mac_header(skb);
@@ -413,6 +414,8 @@ addif(struct aoetgt *t, struct net_device *nd)
 	p = getif(t, NULL);
 	if (!p)
 		return NULL;
+
+	dev_hold(nd);
 	p->nd = nd;
 	p->maxbcnt = DEFAULTBCNT;
 	p->lost = 0;
@@ -424,12 +427,30 @@ static void
 ejectif(struct aoetgt *t, struct aoeif *ifp)
 {
 	struct aoeif *e;
+	struct net_device *nd;
 	ulong n;
 
 	e = t->ifs + NAOEIFS - 1;
+	nd = e->nd;
 	n = (e - ifp) * sizeof *ifp;
 	memmove(ifp, ifp+1, n);
 	e->nd = NULL;
+	dev_put(nd);
+}
+
+void
+aoecmd_flushnet(struct aoedev *d, struct net_device *nd)
+{
+	struct aoetgt **tt, **te;
+	tt = d->targets;
+	te = tt + NTARGETS;
+	for (; tt < te && *tt; tt++) {
+		struct aoetgt *t = *tt;
+		struct aoeif *ifp;
+
+		if ( (ifp = getif(t, nd)) )
+			ejectif(t, ifp);
+	}
 }
 
 static int
drivers/block/aoe/aoedev.c +15 -0
index fa67027..6754db0 100644
@@ -162,6 +162,21 @@ aoedev_flush(const char __user *str, size_t cnt)
 	return 0;
 }
 
+void
+aoedev_ejectnet(struct net_device *nd)
+{
+	struct aoedev *d;
+	unsigned long flags;
+
+	spin_lock_irqsave(&devlist_lock, flags);
+	for (d = devlist; d; d = d->next) {
+		spin_lock(&d->lock);
+		aoecmd_flushnet(d, nd);
+		spin_unlock(&d->lock);
+	}
+	spin_unlock_irqrestore(&d->lock, flags);
+}
+
 /* I'm not really sure that this is a realistic problem, but if the
 network driver goes gonzo let's just leak memory after complaining. */
 static void
drivers/block/aoe/aoemain.c +25 -0
index 7f83ad9..5f6c072 100644
@@ -8,6 +8,8 @@
 #include <linux/blkdev.h>
 #include <linux/module.h>
 #include <linux/skbuff.h>
+#include <linux/notifier.h>
+#include <linux/netdevice.h>
 #include "aoe.h"
 
 MODULE_LICENSE("GPL");
@@ -54,11 +56,29 @@ discover_timer(ulong vp)
 	}
 }
 
+/* Callback on change of state of network device. */
+static int
+aoe_device_event(struct notifier_block *unused,
+		unsigned long event, void *ptr)
+{
+	struct net_device *nd = ptr;
+
+	if (event == NETDEV_UNREGISTER)
+		aoedev_ejectnet(nd);
+
+	return NOTIFY_DONE;
+}
+
+static struct notifier_block aoe_notifier = {
+	.notifier_call = aoe_device_event,
+};
+
 static void
 aoe_exit(void)
 {
 	discover_timer(TKILL);
 
+	unregister_netdevice_notifier(&aoe_notifier);
 	aoenet_exit();
 	unregister_blkdev(AOE_MAJOR, DEVICE_NAME);
 	aoechr_exit();
@@ -83,6 +103,9 @@ aoe_init(void)
 	ret = aoenet_init();
 	if (ret)
 		goto net_fail;
+	ret = register_netdevice_notifier(&aoe_notifier);
+	if (ret)
+		goto notifier_fail;
 	ret = register_blkdev(AOE_MAJOR, DEVICE_NAME);
 	if (ret < 0) {
 		printk(KERN_ERR "aoe: can't register major\n");
@@ -94,6 +117,8 @@ aoe_init(void)
 	return 0;
 
  blkreg_fail:
+	unregister_netdevice_notifier(&aoe_notifier);
+ notifier_fail:
 	aoenet_exit();
  net_fail:
 	aoeblk_exit();
drivers/block/aoe/aoenet.c +4 -1
index ce0d62c..cc3872c 100644
@@ -90,7 +90,10 @@ aoenet_xmit(struct sk_buff_head *queue)
 
 	skb_queue_walk_safe(queue, skb, tmp) {
 		__skb_unlink(skb, queue);
-		dev_queue_xmit(skb);
+		if (skb->dev)
+			dev_queue_xmit(skb);
+		else
+			dev_kfree_skb(skb);
 	}
 }
 
#24 lore raw

Re: [PATCH 04/10] AOE: use rcu to find network device

Quoted 8 lines · On Fri, 13 Nov 2009 16:39:49 -0500 Ed Cashin <[email protected]> wrote:

I will be testing the patch below on Monday. It is based on Stephen Hemminger's starter patch (Thanks!). It adds handling for the case where all of the local network interfaces have become unusable. In that case, sent packets and retransmitted packets are effectively dropped, and the driver's handling of ethernet's unreliability takes care of things as usual. If a local interface becomes usable in time no I/O will fail.

Quoted 4 lines

After testing I'll make a proper changelog entry. In brief this patch allows the aoe driver to correctly take references on the net_device's that it uses, handle their removal via a callback, and handle their complete absence, too.

Looks good, good luck with testing.

#25 lore raw

PATCH 5/10 parisc: use RCU to find network device

In reply to Stephen Hemminger

Another place where RCU can be used instead of read_lock(&dev_base_lock) This is by inspection, don't have platform or cross-build environment to validate.

drivers/parisc/led.c +1 -3
@@ -354,9 +354,8 @@ static __inline__ int led_get_net_activi
 	
 	/* we are running as a workqueue task, so locking dev_base 
 	 * for reading should be OK */
-	read_lock(&dev_base_lock);
 	rcu_read_lock();
-	for_each_netdev(&init_net, dev) {
+	for_each_netdev_rcu(&init_net, dev) {
 	    const struct net_device_stats *stats;
 	    struct in_device *in_dev = __in_dev_get_rcu(dev);
 	    if (!in_dev || !in_dev->ifa_list)
@@ -368,7 +367,6 @@ static __inline__ int led_get_net_activi
 	    tx_total += stats->tx_packets;
 	}
 	rcu_read_unlock();
-	read_unlock(&dev_base_lock);
 
 	retval = 0;
 
#28 lore raw

PATCH 6/10 s390: use RCU to walk list of network devices

In reply to Stephen Hemminger

This is similar to other cases where for_each_netdev_rcu can be used when gathering information.

By inspection, don't have platform or cross-build environment to validate.

arch/s390/appldata/appldata_net_sum.c +5 -3
@@ -83,8 +83,9 @@ static void appldata_get_net_sum_data(vo
 	rx_dropped = 0;
 	tx_dropped = 0;
 	collisions = 0;
-	read_lock(&dev_base_lock);
-	for_each_netdev(&init_net, dev) {
+
+	rcu_read_lock();
+	for_each_netdev_rcu(&init_net, dev) {
 		const struct net_device_stats *stats = dev_get_stats(dev);
 
 		rx_packets += stats->rx_packets;
@@ -98,7 +99,8 @@ static void appldata_get_net_sum_data(vo
 		collisions += stats->collisions;
 		i++;
 	}
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
+
 	net_data->nr_interfaces = i;
 	net_data->rx_packets = rx_packets;
 	net_data->tx_packets = tx_packets;
#29 lore raw

Re: [PATCH 06/10] s390: use RCU to walk list of network devices

Quoted 25 lines · Stephen Hemminger

This is similar to other cases where for_each_netdev_rcu can be used when gathering information.

By inspection, don't have platform or cross-build environment to validate.

arch/s390/appldata/appldata_net_sum.c +3 -2
@@ -83,8 +83,9 @@ static void appldata_get_net_sum_data(vo
 	rx_dropped = 0;
 	tx_dropped = 0;
 	collisions = 0;
-	read_lock(&dev_base_lock);
-	for_each_netdev(&init_net, dev) {
+
+	rcu_read_lock();
+	for_each_netdev_rcu(&init_net, dev) {
 		const struct net_device_stats *stats = dev_get_stats(dev);
 
 		rx_packets += stats->rx_packets;

arch/s390/appldata/appldata_net_sum.c static void appldata_get_net_sum_data(vo · line 106

Quoted 9 lines
arch/s390/appldata/appldata_net_sum.c +2 -1
@@ -98,7 +99,8 @@ static void appldata_get_net_sum_data(vo
 		collisions += stats->collisions;
 		i++;
 	}
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
+
 	net_data->nr_interfaces = i;
 	net_data->rx_packets = rx_packets;
 	net_data->tx_packets = tx_packets;

Not sure if dev_get_stats(dev) could sleep on some devices...

#30 lore raw

Re: [PATCH 06/10] s390: use RCU to walk list of network devices

Quoted 23 lines · On Tue, 10 Nov 2009 19:27:11 +0100 Eric Dumazet <[email protected]> wrote:
Quoted 22 lines · Stephen Hemminger

This is similar to other cases where for_each_netdev_rcu can be used when gathering information.

By inspection, don't have platform or cross-build environment to validate.

arch/s390/appldata/appldata_net_sum.c +3 -2
@@ -83,8 +83,9 @@ static void appldata_get_net_sum_data(vo
 	rx_dropped = 0;
 	tx_dropped = 0;
 	collisions = 0;
-	read_lock(&dev_base_lock);
-	for_each_netdev(&init_net, dev) {
+
+	rcu_read_lock();
+	for_each_netdev_rcu(&init_net, dev) {
 		const struct net_device_stats *stats = dev_get_stats(dev);
 
Quoted 15 lines

arch/s390/appldata/appldata_net_sum.c static void appldata_get_net_sum_data(vo · line 106

Quoted 12 lines
arch/s390/appldata/appldata_net_sum.c +2 -1
 		rx_packets += stats->rx_packets;
@@ -98,7 +99,8 @@ static void appldata_get_net_sum_data(vo
 		collisions += stats->collisions;
 		i++;
 	}
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
+
 	net_data->nr_interfaces = i;
 	net_data->rx_packets = rx_packets;
 	net_data->tx_packets = tx_packets;

Not sure if dev_get_stats(dev) could sleep on some devices...

It would have already been broken since dev_get_stats is previously called with read_lock(), and sleeping with any lock held causes warning.

maintainer Applied
#32 lore raw

Re: [PATCH 06/10] s390: use RCU to walk list of network devices

From: Eric Dumazet <[email protected]>
Date: Tue, 10 Nov 2009 19:40:42 +0100

Quoted 4 lines
Quoted 3 lines · Stephen Hemminger

This is similar to other cases where for_each_netdev_rcu can be used when gathering information.

Quoted 8 lines
Quoted 6 lines

By inspection, don't have platform or cross-build environment to validate.

Applied.

#33 lore raw

PATCH 7/10 decnet: use RCU to find network devices

In reply to Stephen Hemminger

When showing device statistics use RCU rather than read_lock(&dev_base_lock) Compile tested only.

net/decnet/dn_dev.c +6 -8
@@ -856,9 +856,7 @@ int dn_dev_bind_default(__le16 *addr)
 	dev = dn_dev_get_default();
 last_chance:
 	if (dev) {
-		read_lock(&dev_base_lock);
 		rv = dn_dev_get_first(dev, addr);
-		read_unlock(&dev_base_lock);
 		dev_put(dev);
 		if (rv == 0 || dev == init_net.loopback_dev)
 			return rv;
@@ -1323,18 +1321,18 @@ static inline int is_dn_dev(struct net_d
 }
 
 static void *dn_dev_seq_start(struct seq_file *seq, loff_t *pos)
-	__acquires(&dev_base_lock)
+	__acquires(rcu)
 {
 	int i;
 	struct net_device *dev;
 
-	read_lock(&dev_base_lock);
+	rcu_read_lock();
 
 	if (*pos == 0)
 		return SEQ_START_TOKEN;
 
 	i = 1;
-	for_each_netdev(&init_net, dev) {
+	for_each_netdev_rcu(&init_net, dev) {
 		if (!is_dn_dev(dev))
 			continue;
 
@@ -1355,7 +1353,7 @@ static void *dn_dev_seq_next(struct seq_
 	if (v == SEQ_START_TOKEN)
 		dev = net_device_entry(&init_net.dev_base_head);
 
-	for_each_netdev_continue(&init_net, dev) {
+	for_each_netdev_continue_rcu(&init_net, dev) {
 		if (!is_dn_dev(dev))
 			continue;
 
@@ -1366,9 +1364,9 @@ static void *dn_dev_seq_next(struct seq_
 }
 
 static void dn_dev_seq_stop(struct seq_file *seq, void *v)
-	__releases(&dev_base_lock)
+	__releases(rcu)
 {
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 }
 
 static char *dn_type2asc(char type)
#34 lore raw

Re: [PATCH 07/10] decnet: use RCU to find network devices

net/decnet/dn_dev.c static inline int is_dn_dev(struct net_d

Quoted 18 lines · Stephen Hemminger

When showing device statistics use RCU rather than read_lock(&dev_base_lock) Compile tested only.

net/decnet/dn_dev.c +0 -2
@@ -856,9 +856,7 @@ int dn_dev_bind_default(__le16 *addr)
 	dev = dn_dev_get_default();
 last_chance:
 	if (dev) {
-		read_lock(&dev_base_lock);
 		rv = dn_dev_get_first(dev, addr);
-		read_unlock(&dev_base_lock);
 		dev_put(dev);
 		if (rv == 0 || dev == init_net.loopback_dev)
 			return rv;
@@ -1323,18 +1321,18 @@ static inline int is_dn_dev(struct net_d

I dont understand this part. Why previous locking can be avoided ?

#35 lore raw

Re: [PATCH 07/10] decnet: use RCU to find network devices

Quoted 6 lines · On Tue, 10 Nov 2009 19:43:21 +0100 Eric Dumazet <[email protected]> wrote:
Quoted 5 lines · Stephen Hemminger

When showing device statistics use RCU rather than read_lock(&dev_base_lock) Compile tested only.

Quoted 16 lines

net/decnet/dn_dev.c static inline int is_dn_dev(struct net_d

Quoted 13 lines
net/decnet/dn_dev.c +0 -2
@@ -856,9 +856,7 @@ int dn_dev_bind_default(__le16 *addr)
 	dev = dn_dev_get_default();
 last_chance:
 	if (dev) {
-		read_lock(&dev_base_lock);
 		rv = dn_dev_get_first(dev, addr);
-		read_unlock(&dev_base_lock);
 		dev_put(dev);
 		if (rv == 0 || dev == init_net.loopback_dev)
 			return rv;
@@ -1323,18 +1321,18 @@ static inline int is_dn_dev(struct net_d

I dont understand this part. Why previous locking can be avoided ?

dn_dev_get_default acquires a reference on dev so the device can not go away.

It could be the original author meant to ensure the address list doesn't change. If so, then rtnl_lock() should have been used.

#36 lore raw

Re: [PATCH 07/10] decnet: use RCU to find network devices

Quoted 25 lines · Stephen Hemminger
Quoted 6 lines · On Tue, 10 Nov 2009 19:43:21 +0100 Eric Dumazet <[email protected]> wrote:
Quoted 5 lines · Stephen Hemminger

When showing device statistics use RCU rather than read_lock(&dev_base_lock) Compile tested only.

Quoted 15 lines

net/decnet/dn_dev.c static inline int is_dn_dev(struct net_d

Quoted 13 lines
net/decnet/dn_dev.c +0 -2
@@ -856,9 +856,7 @@ int dn_dev_bind_default(__le16 *addr)
 	dev = dn_dev_get_default();
 last_chance:
 	if (dev) {
-		read_lock(&dev_base_lock);
 		rv = dn_dev_get_first(dev, addr);
-		read_unlock(&dev_base_lock);
 		dev_put(dev);
 		if (rv == 0 || dev == init_net.loopback_dev)
 			return rv;
@@ -1323,18 +1321,18 @@ static inline int is_dn_dev(struct net_d

I dont understand this part. Why previous locking can be avoided ?

Quoted 5 lines

dn_dev_get_default acquires a reference on dev so the device can not go away.

It could be the original author meant to ensure the address list doesn't change. If so, then rtnl_lock() should have been used.

Hmm... I spot some bugs during my previous round of patches, so maybe this is a real bug... We should double check.

#37 lore raw

Re: [PATCH 07/10] decnet: use RCU to find network devices

In reply to Stephen Hemminger

Hi,

Quoted 26 lines · Stephen Hemminger
Quoted 6 lines · On Tue, 10 Nov 2009 19:43:21 +0100 Eric Dumazet <[email protected]> wrote:
Quoted 5 lines · Stephen Hemminger

When showing device statistics use RCU rather than read_lock(&dev_base_lock) Compile tested only.

Quoted 16 lines

net/decnet/dn_dev.c static inline int is_dn_dev(struct net_d

Quoted 13 lines
net/decnet/dn_dev.c +0 -2
@@ -856,9 +856,7 @@ int dn_dev_bind_default(__le16 *addr)
 	dev = dn_dev_get_default();
 last_chance:
 	if (dev) {
-		read_lock(&dev_base_lock);
 		rv = dn_dev_get_first(dev, addr);
-		read_unlock(&dev_base_lock);
 		dev_put(dev);
 		if (rv == 0 || dev == init_net.loopback_dev)
 			return rv;
@@ -1323,18 +1321,18 @@ static inline int is_dn_dev(struct net_d

I dont understand this part. Why previous locking can be avoided ?

Quoted 5 lines

dn_dev_get_default acquires a reference on dev so the device can not go away.

It could be the original author meant to ensure the address list doesn't change. If so, then rtnl_lock() should have been used.

The original author has tried to remember what he was thinking when he wrote this code :-)

I think you are right that it should be rtnl_lock() as we don't want the address list changing at this point. On the other hand I notice also that other bits of the code seem to be using dev_base_lock too.

Maybe this is a hang over from another era? I'll have to refresh my memory some more before I can give a verdict on that I'm afraid,

Steve.

#38 lore raw

PATCH 1/2 decnet: add RTNL lock when reading address list

Add missing locking in the case of auto binding to the default device. The address list might change while this code is looking at the list.

Compile tested only, I am not a decnet user.

net/decnet/dn_dev.c +4 -0
@@ -828,13 +828,17 @@ static int dn_dev_get_first(struct net_d
 	struct dn_dev *dn_db = (struct dn_dev *)dev->dn_ptr;
 	struct dn_ifaddr *ifa;
 	int rv = -ENODEV;
+
 	if (dn_db == NULL)
 		goto out;
+
+	rtnl_lock();
 	ifa = dn_db->ifa_list;
 	if (ifa != NULL) {
 		*addr = ifa->ifa_local;
 		rv = 0;
 	}
+	rtnl_unlock();
 out:
 	return rv;
 }
#39 lore raw

PATCH 2/2 decnet: convert dndev_lock to spinlock

There is no reason for this lock to be reader/writer since the reader only has lock held for a very brief period. The overhead of read_lock is more expensive than spinlock.

Compile tested only, I am not a decnet user.

net/decnet/dn_dev.c +12 -7
@@ -68,7 +68,7 @@ extern struct neigh_table dn_neigh_table
  */
 __le16 decnet_address = 0;
 
-static DEFINE_RWLOCK(dndev_lock);
+static DEFINE_SPINLOCK(dndev_lock);
 static struct net_device *decnet_default_device;
 static BLOCKING_NOTIFIER_HEAD(dnaddr_chain);
 
@@ -557,7 +557,8 @@ rarok:
 struct net_device *dn_dev_get_default(void)
 {
 	struct net_device *dev;
-	read_lock(&dndev_lock);
+
+	spin_lock(&dndev_lock);
 	dev = decnet_default_device;
 	if (dev) {
 		if (dev->dn_ptr)
@@ -565,7 +566,8 @@ struct net_device *dn_dev_get_default(vo
 		else
 			dev = NULL;
 	}
-	read_unlock(&dndev_lock);
+	spin_unlock(&dndev_lock);
+
 	return dev;
 }
 
@@ -575,13 +577,15 @@ int dn_dev_set_default(struct net_device
 	int rv = -EBUSY;
 	if (!dev->dn_ptr)
 		return -ENODEV;
-	write_lock(&dndev_lock);
+
+	spin_lock(&dndev_lock);
 	if (force || decnet_default_device == NULL) {
 		old = decnet_default_device;
 		decnet_default_device = dev;
 		rv = 0;
 	}
-	write_unlock(&dndev_lock);
+	spin_unlock(&dndev_lock);
+
 	if (old)
 		dev_put(old);
 	return rv;
@@ -589,13 +593,14 @@ int dn_dev_set_default(struct net_device
 
 static void dn_dev_check_default(struct net_device *dev)
 {
-	write_lock(&dndev_lock);
+	spin_lock(&dndev_lock);
 	if (dev == decnet_default_device) {
 		decnet_default_device = NULL;
 	} else {
 		dev = NULL;
 	}
-	write_unlock(&dndev_lock);
+	spin_unlock(&dndev_lock);
+
 	if (dev)
 		dev_put(dev);
 }
maintainer Applied
#40 lore raw

Re: [PATCH 2/2] decnet: convert dndev_lock to spinlock

From: Stephen Hemminger <[email protected]>
Date: Wed, 11 Nov 2009 09:40:36 -0800

Quoted 7 lines

There is no reason for this lock to be reader/writer since the reader only has lock held for a very brief period. The overhead of read_lock is more expensive than spinlock.

Compile tested only, I am not a decnet user.

Applied to net-next-2.6

maintainer Applied
#41 lore raw

Re: [PATCH 1/2] decnet: add RTNL lock when reading address list

In reply to Stephen Hemminger

From: Stephen Hemminger <[email protected]>
Date: Wed, 11 Nov 2009 09:39:27 -0800

Quoted 7 lines

Add missing locking in the case of auto binding to the default device. The address list might change while this code is looking at the list.

Compile tested only, I am not a decnet user.

Applied to net-next-2.6

#43 lore raw

PATCH 8/10 ipv6: use RCU to walk list of network devices

In reply to Stephen Hemminger

No longer need read_lock(&dev_base_lock), use RCU instead. This also needs to be optimized for large number of devices.

net/ipv6/addrconf.c +3 -3
@@ -3828,9 +3828,9 @@ static int inet6_dump_ifinfo(struct sk_b
 	struct net_device *dev;
 	struct inet6_dev *idev;
 
-	read_lock(&dev_base_lock);
+	rcu_read_lock();
 	idx = 0;
-	for_each_netdev(net, dev) {
+	for_each_netdev_rcu(net, dev) {
 		if (idx < s_idx)
 			goto cont;
 		if ((idev = in6_dev_get(dev)) == NULL)
@@ -3843,7 +3843,7 @@ static int inet6_dump_ifinfo(struct sk_b
 cont:
 		idx++;
 	}
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 	cb->args[0] = idx;
 
 	return skb->len;
net/ipv6/anycast.c +5 -5
@@ -431,7 +431,7 @@ static inline struct ifacaddr6 *ac6_get_
 	struct net *net = seq_file_net(seq);
 
 	state->idev = NULL;
-	for_each_netdev(net, state->dev) {
+	for_each_netdev_rcu(net, state->dev) {
 		struct inet6_dev *idev;
 		idev = in6_dev_get(state->dev);
 		if (!idev)
@@ -482,9 +482,9 @@ static struct ifacaddr6 *ac6_get_idx(str
 }
 
 static void *ac6_seq_start(struct seq_file *seq, loff_t *pos)
-	__acquires(dev_base_lock)
+	__acquires(rcu)
 {
-	read_lock(&dev_base_lock);
+	rcu_read_lock();
 	return ac6_get_idx(seq, *pos);
 }
 
@@ -497,14 +497,14 @@ static void *ac6_seq_next(struct seq_fil
 }
 
 static void ac6_seq_stop(struct seq_file *seq, void *v)
-	__releases(dev_base_lock)
+	__releases(rcu)
 {
 	struct ac6_iter_state *state = ac6_seq_private(seq);
 	if (likely(state->idev != NULL)) {
 		read_unlock_bh(&state->idev->lock);
 		in6_dev_put(state->idev);
 	}
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 }
 
 static int ac6_seq_show(struct seq_file *seq, void *v)
net/ipv6/mcast.c +10 -10
@@ -2375,7 +2375,7 @@ static inline struct ifmcaddr6 *igmp6_mc
 	struct net *net = seq_file_net(seq);
 
 	state->idev = NULL;
-	for_each_netdev(net, state->dev) {
+	for_each_netdev_rcu(net, state->dev) {
 		struct inet6_dev *idev;
 		idev = in6_dev_get(state->dev);
 		if (!idev)
@@ -2426,9 +2426,9 @@ static struct ifmcaddr6 *igmp6_mc_get_id
 }
 
 static void *igmp6_mc_seq_start(struct seq_file *seq, loff_t *pos)
-	__acquires(dev_base_lock)
+	__acquires(rcu)
 {
-	read_lock(&dev_base_lock);
+	rcu_read_lock();
 	return igmp6_mc_get_idx(seq, *pos);
 }
 
@@ -2441,7 +2441,7 @@ static void *igmp6_mc_seq_next(struct se
 }
 
 static void igmp6_mc_seq_stop(struct seq_file *seq, void *v)
-	__releases(dev_base_lock)
+	__releases(rcu)
 {
 	struct igmp6_mc_iter_state *state = igmp6_mc_seq_private(seq);
 	if (likely(state->idev != NULL)) {
@@ -2450,7 +2450,7 @@ static void igmp6_mc_seq_stop(struct seq
 		state->idev = NULL;
 	}
 	state->dev = NULL;
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 }
 
 static int igmp6_mc_seq_show(struct seq_file *seq, void *v)
@@ -2507,7 +2507,7 @@ static inline struct ip6_sf_list *igmp6_
 
 	state->idev = NULL;
 	state->im = NULL;
-	for_each_netdev(net, state->dev) {
+	for_each_netdev_rcu(net, state->dev) {
 		struct inet6_dev *idev;
 		idev = in6_dev_get(state->dev);
 		if (unlikely(idev == NULL))
@@ -2573,9 +2573,9 @@ static struct ip6_sf_list *igmp6_mcf_get
 }
 
 static void *igmp6_mcf_seq_start(struct seq_file *seq, loff_t *pos)
-	__acquires(dev_base_lock)
+	__acquires(rcu)
 {
-	read_lock(&dev_base_lock);
+	rcu_read_lock();
 	return *pos ? igmp6_mcf_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
 }
 
@@ -2591,7 +2591,7 @@ static void *igmp6_mcf_seq_next(struct s
 }
 
 static void igmp6_mcf_seq_stop(struct seq_file *seq, void *v)
-	__releases(dev_base_lock)
+	__releases(rcu)
 {
 	struct igmp6_mcf_iter_state *state = igmp6_mcf_seq_private(seq);
 	if (likely(state->im != NULL)) {
@@ -2604,7 +2604,7 @@ static void igmp6_mcf_seq_stop(struct se
 		state->idev = NULL;
 	}
 	state->dev = NULL;
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 }
 
 static int igmp6_mcf_seq_show(struct seq_file *seq, void *v)
maintainer
#44 lore raw

Re: [PATCH 08/10] ipv6: use RCU to walk list of network devices

From: Stephen Hemminger <[email protected]>
Date: Tue, 10 Nov 2009 09:54:54 -0800

Quoted 4 lines

No longer need read_lock(&dev_base_lock), use RCU instead. This also needs to be optimized for large number of devices.

As Eric mentioned, superceded by his patch.

#45 lore raw

PATCH 1/1 ipv6: use RCU to walk list of network devices

In reply to Stephen Hemminger
Quoted 4 lines · Stephen Hemminger

No longer need read_lock(&dev_base_lock), use RCU instead. This also needs to be optimized for large number of devices.

I gave a look at your patch Stephen and found we need a new next_net_device_rcu(struct net_device *dev) as well, as next_net_device() is not RCU safe.

(followup patch is probably needed to use it in net/ipv4/igmp.c, after commit 61fbab77a843d2e77232 : IPV4: use rcu to walk list of devices in IGMP)

We also can avoid taking references on inet6_dev structs.

[PATCH net-next-2.6] ipv6: use RCU to walk list of network devices

No longer need read_lock(&dev_base_lock), use RCU instead. We also can avoid taking references on inet6_dev structs.

---
 include/linux/netdevice.h |   10 +++++++
 net/ipv6/anycast.c        |   29 +++++++++-----------
 net/ipv6/mcast.c          |   51 ++++++++++++++++--------------------
 3 files changed, 47 insertions(+), 43 deletions(-)
include/linux/netdevice.h +10 -0
index 083b598..2734a67 100644
@@ -1093,6 +1093,16 @@ static inline struct net_device *next_net_device(struct net_device *dev)
 	return lh == &net->dev_base_head ? NULL : net_device_entry(lh);
 }
 
+static inline struct net_device *next_net_device_rcu(struct net_device *dev)
+{
+	struct list_head *lh;
+	struct net *net;
+
+	net = dev_net(dev);
+	lh = rcu_dereference(dev->dev_list.next);
+	return lh == &net->dev_base_head ? NULL : net_device_entry(lh);
+}
+
 static inline struct net_device *first_net_device(struct net *net)
 {
 	return list_empty(&net->dev_base_head) ? NULL :
net/ipv6/anycast.c +14 -15
index 2f00ca8..f1c74c8 100644
@@ -431,9 +431,9 @@ static inline struct ifacaddr6 *ac6_get_first(struct seq_file *seq)
 	struct net *net = seq_file_net(seq);
 
 	state->idev = NULL;
-	for_each_netdev(net, state->dev) {
+	for_each_netdev_rcu(net, state->dev) {
 		struct inet6_dev *idev;
-		idev = in6_dev_get(state->dev);
+		idev = __in6_dev_get(state->dev);
 		if (!idev)
 			continue;
 		read_lock_bh(&idev->lock);
@@ -443,7 +443,6 @@ static inline struct ifacaddr6 *ac6_get_first(struct seq_file *seq)
 			break;
 		}
 		read_unlock_bh(&idev->lock);
-		in6_dev_put(idev);
 	}
 	return im;
 }
@@ -454,16 +453,15 @@ static struct ifacaddr6 *ac6_get_next(struct seq_file *seq, struct ifacaddr6 *im
 
 	im = im->aca_next;
 	while (!im) {
-		if (likely(state->idev != NULL)) {
+		if (likely(state->idev != NULL))
 			read_unlock_bh(&state->idev->lock);
-			in6_dev_put(state->idev);
-		}
-		state->dev = next_net_device(state->dev);
+
+		state->dev = next_net_device_rcu(state->dev);
 		if (!state->dev) {
 			state->idev = NULL;
 			break;
 		}
-		state->idev = in6_dev_get(state->dev);
+		state->idev = __in6_dev_get(state->dev);
 		if (!state->idev)
 			continue;
 		read_lock_bh(&state->idev->lock);
@@ -482,29 +480,30 @@ static struct ifacaddr6 *ac6_get_idx(struct seq_file *seq, loff_t pos)
 }
 
 static void *ac6_seq_start(struct seq_file *seq, loff_t *pos)
-	__acquires(dev_base_lock)
+	__acquires(RCU)
 {
-	read_lock(&dev_base_lock);
+	rcu_read_lock();
 	return ac6_get_idx(seq, *pos);
 }
 
 static void *ac6_seq_next(struct seq_file *seq, void *v, loff_t *pos)
 {
-	struct ifacaddr6 *im;
-	im = ac6_get_next(seq, v);
+	struct ifacaddr6 *im = ac6_get_next(seq, v);
+
 	++*pos;
 	return im;
 }
 
 static void ac6_seq_stop(struct seq_file *seq, void *v)
-	__releases(dev_base_lock)
+	__releases(RCU)
 {
 	struct ac6_iter_state *state = ac6_seq_private(seq);
+
 	if (likely(state->idev != NULL)) {
 		read_unlock_bh(&state->idev->lock);
-		in6_dev_put(state->idev);
+		state->idev = NULL;
 	}
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 }
 
 static int ac6_seq_show(struct seq_file *seq, void *v)
net/ipv6/mcast.c +23 -28

145 lines not on this page · raw message

maintainer Applied
#46 lore raw

Re: [PATCH net-next-2.6] ipv6: use RCU to walk list of network devices

From: Eric Dumazet <[email protected]>
Date: Thu, 12 Nov 2009 04:34:30 +0100

Quoted 6 lines

[PATCH net-next-2.6] ipv6: use RCU to walk list of network devices

No longer need read_lock(&dev_base_lock), use RCU instead. We also can avoid taking references on inet6_dev structs.

Applied.

#47 lore raw

PATCH 9/10 IPV4: use rcu to walk list of devices in IGMP

In reply to Stephen Hemminger

This also needs to be optimized for large number of devices.

net/ipv4/igmp.c +10 -8
@@ -2311,7 +2311,7 @@ static inline struct ip_mc_list *igmp_mc
 	struct igmp_mc_iter_state *state = igmp_mc_seq_private(seq);
 
 	state->in_dev = NULL;
-	for_each_netdev(net, state->dev) {
+	for_each_netdev_rcu(net, state->dev) {
 		struct in_device *in_dev;
 		in_dev = in_dev_get(state->dev);
 		if (!in_dev)
@@ -2361,9 +2361,9 @@ static struct ip_mc_list *igmp_mc_get_id
 }
 
 static void *igmp_mc_seq_start(struct seq_file *seq, loff_t *pos)
-	__acquires(dev_base_lock)
+	__acquires(rcu)
 {
-	read_lock(&dev_base_lock);
+	rcu_read_lock();
 	return *pos ? igmp_mc_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
 }
 
@@ -2379,7 +2379,7 @@ static void *igmp_mc_seq_next(struct seq
 }
 
 static void igmp_mc_seq_stop(struct seq_file *seq, void *v)
-	__releases(dev_base_lock)
+	__releases(rcu)
 {
 	struct igmp_mc_iter_state *state = igmp_mc_seq_private(seq);
 	if (likely(state->in_dev != NULL)) {
@@ -2388,7 +2388,7 @@ static void igmp_mc_seq_stop(struct seq_
 		state->in_dev = NULL;
 	}
 	state->dev = NULL;
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 }
 
 static int igmp_mc_seq_show(struct seq_file *seq, void *v)
@@ -2462,7 +2462,7 @@ static inline struct ip_sf_list *igmp_mc
 
 	state->idev = NULL;
 	state->im = NULL;
-	for_each_netdev(net, state->dev) {
+	for_each_netdev_rcu(net, state->dev) {
 		struct in_device *idev;
 		idev = in_dev_get(state->dev);
 		if (unlikely(idev == NULL))
@@ -2528,8 +2528,9 @@ static struct ip_sf_list *igmp_mcf_get_i
 }
 
 static void *igmp_mcf_seq_start(struct seq_file *seq, loff_t *pos)
+	__acquires(rcu)
 {
-	read_lock(&dev_base_lock);
+	rcu_read_lock();
 	return *pos ? igmp_mcf_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
 }
 
@@ -2545,6 +2546,7 @@ static void *igmp_mcf_seq_next(struct se
 }
 
 static void igmp_mcf_seq_stop(struct seq_file *seq, void *v)
+	__releases(rcu)
 {
 	struct igmp_mcf_iter_state *state = igmp_mcf_seq_private(seq);
 	if (likely(state->im != NULL)) {
@@ -2557,7 +2559,7 @@ static void igmp_mcf_seq_stop(struct seq
 		state->idev = NULL;
 	}
 	state->dev = NULL;
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 }
 
 static int igmp_mcf_seq_show(struct seq_file *seq, void *v)
Acked-by on v1 9/10
#48 lore raw

Re: [PATCH 09/10] IPV4: use rcu to walk list of devices in IGMP

Review

Quoted 7 lines · Stephen Hemminger

This also needs to be optimized for large number of devices.

static void igmp_mcf_seq_stop(struct seq_file *seq, void *v)

(quoted diff) +1 -0
+	__releases(rcu)

Minor note : we usually use RCU instead of rcu, but it doesnt really matter.

#50 lore raw

PATCH 10/10 CAN: use dev_get_by_index_rcu

In reply to Stephen Hemminger

Use new function to avoid doing read_lock().

net/can/bcm.c +3 -3
@@ -139,13 +139,13 @@ static char *bcm_proc_getifname(char *re
 	if (!ifindex)
 		return "any";
 
-	read_lock(&dev_base_lock);
-	dev = __dev_get_by_index(&init_net, ifindex);
+	rcu_read_lock();
+	dev = dev_get_by_index_rcu(&init_net, ifindex);
 	if (dev)
 		strcpy(result, dev->name);
 	else
 		strcpy(result, "???");
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 
 	return result;
 }
Acked-by on v1 10/10
#51 lore raw

Re: [PATCH 10/10] CAN: use dev_get_by_index_rcu

Quoted 14 lines · Stephen Hemminger

Use new function to avoid doing read_lock().

net/can/bcm.c +1 -2
@@ -139,13 +139,13 @@ static char *bcm_proc_getifname(char *re
 	if (!ifindex)
 		return "any";
 
-	read_lock(&dev_base_lock);
-	dev = __dev_get_by_index(&init_net, ifindex);
+	rcu_read_lock();

net/can/bcm.c static char *bcm_proc_getifname(char *re · line 151

Quoted 10 lines
net/can/bcm.c +2 -1
+	dev = dev_get_by_index_rcu(&init_net, ifindex);
 	if (dev)
 		strcpy(result, dev->name);
 	else
 		strcpy(result, "???");
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 
 	return result;
 }

I was pretty sure I had already done this one...

Ah yes, that was planned after a bugfix for net-2.6.

#52 lore raw

Re: [PATCH 10/10] CAN: use dev_get_by_index_rcu

Quoted 28 lines · Eric Dumazet
Quoted 14 lines · Stephen Hemminger

Use new function to avoid doing read_lock().

net/can/bcm.c +1 -2
@@ -139,13 +139,13 @@ static char *bcm_proc_getifname(char *re
 	if (!ifindex)
 		return "any";
 
-	read_lock(&dev_base_lock);
-	dev = __dev_get_by_index(&init_net, ifindex);
+	rcu_read_lock();

net/can/bcm.c static char *bcm_proc_getifname(char *re · line 151

Quoted 10 lines
net/can/bcm.c +2 -1
+	dev = dev_get_by_index_rcu(&init_net, ifindex);
 	if (dev)
 		strcpy(result, dev->name);
 	else
 		strcpy(result, "???");
-	read_unlock(&dev_base_lock);
+	rcu_read_unlock();
 
 	return result;
 }

I was pretty sure I had already done this one...

Quoted 4 lines

Ah yes, that was planned after a bugfix for net-2.6.

Thanks everyone!

maintainer Applied
#53 lore raw

Re: [PATCH 10/10] CAN: use dev_get_by_index_rcu

From: Oliver Hartkopp <[email protected]>
Date: Wed, 11 Nov 2009 06:54:24 +0100

Quoted 5 lines
Quoted 4 lines · Eric Dumazet
Quoted 3 lines · Stephen Hemminger

Use new function to avoid doing read_lock().

...

Quoted 6 lines
Quoted 2 lines

Thanks everyone!

Applied.

#54 lore raw
In reply to Stephen Hemminger
Quoted 4 lines · Stephen Hemminger

The goal is to eliminate dev_base_lock completely, and just use RCU and rtnl_mutex for network devices. This series gets rid of the many of the users of dev_base_lock just for reading.

Nice, but I was doing all this work Stephen... maybe I am too slow ?

I believe you missed one of my patch (but David is traveling)

[PATCH net-next-2.6] ipv6: speedup inet6_dump_ifinfo()

This conflicts with your :

[PATCH 08/10] ipv6: use RCU to walk list of network devices

#55 lore raw
Quoted 8 lines · On Tue, 10 Nov 2009 19:18:14 +0100 Eric Dumazet <[email protected]> wrote:
Quoted 4 lines · Stephen Hemminger

The goal is to eliminate dev_base_lock completely, and just use RCU and rtnl_mutex for network devices. This series gets rid of the many of the users of dev_base_lock just for reading.

Nice, but I was doing all this work Stephen... maybe I am too slow ?

Quoted 8 lines

I believe you missed one of my patch (but David is traveling)

[PATCH net-next-2.6] ipv6: speedup inet6_dump_ifinfo()

This conflicts with your :

[PATCH 08/10] ipv6: use RCU to walk list of network devices

Inflight conflict.

#56 lore raw
In reply to Eric Dumazet

I was just trying to pick up the stragglers you left behind :-)

The nasty cases left are bonding (whose existing locking model is a pile of crap), and sysfs (slightly less stinky). The bonding code just needs to be rewritten to have a sane/simple model.

#57 lore raw
Quoted 6 lines · Stephen Hemminger

I was just trying to pick up the stragglers you left behind :-)

The nasty cases left are bonding (whose existing locking model is a pile of crap), and sysfs (slightly less stinky). The bonding code just needs to be rewritten to have a sane/simple model.

I was thinking of improving bonding, so it might be good to coordinate our work :)

1) Get rid of rwlock in tx fast path. (RCU ? did I said RCU)

2) multi queue support (Got two dual 82599 dual ports cards from Intel :) )

#58 lore raw
Quoted 9 lines · On Tue, 10 Nov 2009 19:39:20 +0100 Eric Dumazet <[email protected]> wrote:
Quoted 6 lines · Stephen Hemminger

I was just trying to pick up the stragglers you left behind :-)

The nasty cases left are bonding (whose existing locking model is a pile of crap), and sysfs (slightly less stinky). The bonding code just needs to be rewritten to have a sane/simple model.

I was thinking of improving bonding, so it might be good to coordinate our work :)

I'll be happy to review, but was avoiding doing any serious work on it.

Quoted 4 lines

1) Get rid of rwlock in tx fast path. (RCU ? did I said RCU)

2) multi queue support (Got two dual 82599 dual ports cards from Intel :) )

I hope to get back to doing netlink interface for bridging, then bonding over next few weeks.