Skip to contentSkip to conversation

ipv4: factorize cache clearing for batched unregister operations

PATCH Applied Octavian Purdila 4 participants 7 messages Nov 16, 2009 – Nov 18, 2009

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 2/2 ipv4: factorize cache clearing for batched unregister operations

---
 net/ipv4/fib_frontend.c |   11 ++++++-----
 1 files changed, 6 insertions(+), 5 deletions(-)
net/ipv4/fib_frontend.c +6 -5
index 816e218..6c1e56a 100644
@@ -895,11 +895,11 @@ static void nl_fib_lookup_exit(struct net *net)
 	net->ipv4.fibnl = NULL;
 }
 
-static void fib_disable_ip(struct net_device *dev, int force)
+static void fib_disable_ip(struct net_device *dev, int force, int delay)
 {
 	if (fib_sync_down_dev(dev, force))
 		fib_flush(dev_net(dev));
-	rt_cache_flush(dev_net(dev), 0);
+	rt_cache_flush(dev_net(dev), delay);
 	arp_ifdown(dev);
 }
 
@@ -922,7 +922,7 @@ static int fib_inetaddr_event(struct notifier_block *this, unsigned long event,
 			/* Last address was deleted from this interface.
 			   Disable IP.
 			 */
-			fib_disable_ip(dev, 1);
+			fib_disable_ip(dev, 1, 0);
 		} else {
 			rt_cache_flush(dev_net(dev), -1);
 		}
@@ -937,7 +937,7 @@ static int fib_netdev_event(struct notifier_block *this, unsigned long event, vo
 	struct in_device *in_dev = __in_dev_get_rtnl(dev);
 
 	if (event == NETDEV_UNREGISTER) {
-		fib_disable_ip(dev, 2);
+		fib_disable_ip(dev, 2, -1);
 		return NOTIFY_DONE;
 	}
 
@@ -955,10 +955,11 @@ static int fib_netdev_event(struct notifier_block *this, unsigned long event, vo
 		rt_cache_flush(dev_net(dev), -1);
 		break;
 	case NETDEV_DOWN:
-		fib_disable_ip(dev, 0);
+		fib_disable_ip(dev, 0, 0);
 		break;
 	case NETDEV_CHANGEMTU:
 	case NETDEV_CHANGE:
+	case NETDEV_UNREGISTER_PERNET:
 		rt_cache_flush(dev_net(dev), 0);
 		break;
 	}
Signature · 1 line
1.5.6.5
#2 lore raw
Quoted 1 line · Stephen Hemminger

And what about IPV6?

It's not possible to create more than a few thousand IPv6 interfaces in my tests with a 4GB vm. The culprit seems to be the per-cpu arrays for statistics using vmalloc() and wasting gobs of memory as a result (memory usage seems to be 10x per interface vs what IPv4 is using), but I haven't dug too deeply yet.

-ben

maintainer
#3 lore raw

From: Benjamin LaHaise <[email protected]>
Date: Mon, 16 Nov 2009 19:22:38 -0500

Quoted 5 lines

It's not possible to create more than a few thousand IPv6 interfaces in my tests with a 4GB vm. The culprit seems to be the per-cpu arrays for statistics using vmalloc() and wasting gobs of memory as a result (memory usage seems to be 10x per interface vs what IPv4 is using), but I haven't dug too deeply yet.

That definitely should be fixed.

#4 lore raw
In reply to Stephen Hemminger
Quoted 10 lines · On Tuesday 17 November 2009 02:06:46 you wrote:

On Tue, 17 Nov 2009 01:49:49 +0200

Quoted 4 lines · Octavian Purdila
---
 net/ipv4/fib_frontend.c |   11 ++++++-----
 1 files changed, 6 insertions(+), 5 deletions(-)

And what about IPV6?

IPv6 did not appear during profiling my simple dummy test and also a quick scan through the code did not reveal anything obvious on the unregister path which could be factorized.

BTW, it appears we don't have a route cache for IPv6. Right?

I known we will run into other scalability issues when we will set IPv4/IPv6 addresses, and that is one of the next items on my list to profile :)

maintainer
#5 lore raw

From: Octavian Purdila <[email protected]>
Date: Tue, 17 Nov 2009 09:22:08 +0200

Quoted 1 line

BTW, it appears we don't have a route cache for IPv6. Right?

That's right, no IPV6 route cache.

But it doesn't seem to matter so much, as routing tables are still small in that realm. Although I heard the trend is for that to change.

I also suspect that with things like policy routing we could run into problems, but we have a multi-way tree for the IPV6 routing table datastructure explicitly to handle those kinds of things.