Re: [PATCH net v4 01/12] net: core: limit nested device depth

From
Sabrina Dubroca <sd@queasysnail.net>
Date
2019-10-10 10:19:35
Also in
linux-wireless
Reply to
Taehee Yoo, 2019-09-28
2019-09-28, 16:48:32 +0000, Taehee Yoo wrote:
quoted hunk ↗ jump to hunk
@@ -6790,23 +6878,45 @@ int netdev_walk_all_lower_dev(struct net_device *dev,
 					void *data),
 			      void *data)
 {
-	struct net_device *ldev;
-	struct list_head *iter;
-	int ret;
+	struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
+	struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
+	int ret, cur = 0;
 
-	for (iter = &dev->adj_list.lower,
-	     ldev = netdev_next_lower_dev(dev, &iter);
-	     ldev;
-	     ldev = netdev_next_lower_dev(dev, &iter)) {
-		/* first is the lower device itself */
-		ret = fn(ldev, data);
-		if (ret)
-			return ret;
+	now = dev;
+	iter = &dev->adj_list.lower;
 
-		/* then look at all of its lower devices */
-		ret = netdev_walk_all_lower_dev(ldev, fn, data);
-		if (ret)
-			return ret;
+	while (1) {
+		if (now != dev) {
+			ret = fn(now, data);
+			if (ret)
+				return ret;
+		}
+
+		next = NULL;
+		while (1) {
+			ldev = netdev_next_lower_dev(now, &iter);
+			if (!ldev)
+				break;
+
+			if (!next) {
+				next = ldev;
+				niter = &ldev->adj_list.lower;
+			} else {
+				dev_stack[cur] = ldev;
+				iter_stack[cur++] = &ldev->adj_list.lower;
+				break;
+			}
+		}
+
+		if (!next) {
+			if (!cur)
+				return 0;
Hmm, I don't think this condition is correct.

If we have this topology:


                bridge0
                /  |  \
               /   |   \
              /    |    \
        dummy0   vlan1   vlan2
                   |       \
                 dummy1    dummy2

We end up with the expected lower/upper levels for all devices:

    | device  | upper | lower |
    |---------+-------+-------|
    | dummy0  |     2 |     1 |
    | dummy1  |     3 |     1 |
    | dummy2  |     3 |     1 |
    | vlan1   |     2 |     2 |
    | vlan2   |     2 |     2 |
    | bridge0 |     1 |     3 |


If we then add macvlan0 on top of bridge0:


                macvlan0
                   |
                   |
                bridge0
                /  |  \
               /   |   \
              /    |    \
        dummy0   vlan1   vlan2
                   |       \
                 dummy1    dummy2


we can observe that __netdev_update_upper_level is only called for
some of the devices under bridge0. I added a perf probe:

 # perf probe -a '__netdev_update_upper_level dev->name:string'

which gets hit for bridge0 (called directly by
__netdev_upper_dev_link) and then dummy0, vlan1, dummy1. It is never
called for vlan2 and dummy2.

After this, we have the following levels (*):

    | device   | upper | lower |
    |----------+-------+-------|
    | dummy0   |     3 |     1 |
    | dummy1   |     4 |     1 |
    | dummy2   |     3 |     1 |
    | vlan1    |     3 |     2 |
    | vlan2    |     2 |     2 |
    | bridge0  |     2 |     3 |
    | macvlan0 |     1 |     4 |

For dummy0, dummy1, vlan1, the upper level has increased by 1, as
expected. For dummy2 and vlan2, it's still the same, which is wrong.


(*) observed easily by adding another probe:

 # perf probe -a 'dev_get_stats dev->name:string dev->upper_level dev->lower_level'

and running "ip link"

Or you can just add prints and recompile, of course :)
quoted hunk ↗ jump to hunk
+			next = dev_stack[--cur];
+			niter = iter_stack[cur];
+		}
+
+		now = next;
+		iter = niter;
 	}
 
 	return 0;
-- 
Sabrina
Thread (37 messages) 37 messages, 4 authors, last reply 2019-10-21
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help