net: percpu net_device refcount

We tried very hard to remove all possible dev_hold()/dev_put() pairs in
network stack, using RCU conversions.

There is still an unavoidable device refcount change for every dst we
create/destroy, and this can slow down some workloads (routers or some
app servers, mmap af_packet)

We can switch to a percpu refcount implementation, now dynamic per_cpu
infrastructure is mature. On a 64 cpus machine, this consumes 256 bytes
per device.

On x86, dev_hold(dev) code :

before
        lock    incl 0x280(%ebx)
after:
        movl    0x260(%ebx),%eax
        incl    fs:(%eax)

Stress bench :

(Sending 160.000.000 UDP frames,
IP route cache disabled, dual E5540 @2.53GHz,
32bit kernel, FIB_TRIE)

Before:

real    1m1.662s
user    0m14.373s
sys     12m55.960s

After:

real    0m51.179s
user    0m15.329s
sys     10m15.942s

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/net/core/dev.c b/net/core/dev.c
index 193eafa..04972a4 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -5192,9 +5192,6 @@
 	 */
 	dev->reg_state = NETREG_DUMMY;
 
-	/* initialize the ref count */
-	atomic_set(&dev->refcnt, 1);
-
 	/* NAPI wants this */
 	INIT_LIST_HEAD(&dev->napi_list);
 
@@ -5202,6 +5199,11 @@
 	set_bit(__LINK_STATE_PRESENT, &dev->state);
 	set_bit(__LINK_STATE_START, &dev->state);
 
+	/* Note : We dont allocate pcpu_refcnt for dummy devices,
+	 * because users of this 'device' dont need to change
+	 * its refcount.
+	 */
+
 	return 0;
 }
 EXPORT_SYMBOL_GPL(init_dummy_netdev);
@@ -5243,6 +5245,16 @@
 }
 EXPORT_SYMBOL(register_netdev);
 
+int netdev_refcnt_read(const struct net_device *dev)
+{
+	int i, refcnt = 0;
+
+	for_each_possible_cpu(i)
+		refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
+	return refcnt;
+}
+EXPORT_SYMBOL(netdev_refcnt_read);
+
 /*
  * netdev_wait_allrefs - wait until all references are gone.
  *
@@ -5257,11 +5269,14 @@
 static void netdev_wait_allrefs(struct net_device *dev)
 {
 	unsigned long rebroadcast_time, warning_time;
+	int refcnt;
 
 	linkwatch_forget_dev(dev);
 
 	rebroadcast_time = warning_time = jiffies;
-	while (atomic_read(&dev->refcnt) != 0) {
+	refcnt = netdev_refcnt_read(dev);
+
+	while (refcnt != 0) {
 		if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
 			rtnl_lock();
 
@@ -5288,11 +5303,13 @@
 
 		msleep(250);
 
+		refcnt = netdev_refcnt_read(dev);
+
 		if (time_after(jiffies, warning_time + 10 * HZ)) {
 			printk(KERN_EMERG "unregister_netdevice: "
 			       "waiting for %s to become free. Usage "
 			       "count = %d\n",
-			       dev->name, atomic_read(&dev->refcnt));
+			       dev->name, refcnt);
 			warning_time = jiffies;
 		}
 	}
@@ -5350,7 +5367,7 @@
 		netdev_wait_allrefs(dev);
 
 		/* paranoia */
-		BUG_ON(atomic_read(&dev->refcnt));
+		BUG_ON(netdev_refcnt_read(dev));
 		WARN_ON(rcu_dereference_raw(dev->ip_ptr));
 		WARN_ON(dev->ip6_ptr);
 		WARN_ON(dev->dn_ptr);
@@ -5520,9 +5537,13 @@
 	dev = PTR_ALIGN(p, NETDEV_ALIGN);
 	dev->padded = (char *)dev - (char *)p;
 
-	if (dev_addr_init(dev))
+	dev->pcpu_refcnt = alloc_percpu(int);
+	if (!dev->pcpu_refcnt)
 		goto free_tx;
 
+	if (dev_addr_init(dev))
+		goto free_pcpu;
+
 	dev_mc_init(dev);
 	dev_uc_init(dev);
 
@@ -5553,6 +5574,8 @@
 
 free_tx:
 	kfree(tx);
+free_pcpu:
+	free_percpu(dev->pcpu_refcnt);
 free_p:
 	kfree(p);
 	return NULL;
@@ -5586,6 +5609,9 @@
 	list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
 		netif_napi_del(p);
 
+	free_percpu(dev->pcpu_refcnt);
+	dev->pcpu_refcnt = NULL;
+
 	/*  Compatibility with error handling in drivers */
 	if (dev->reg_state == NETREG_UNINITIALIZED) {
 		kfree((char *)dev - dev->padded);