xref: /linux/net/core/link_watch.c (revision 71dfa617ea9f18e4585fe78364217cd32b1fc382)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Linux network device link state notification
4  *
5  * Author:
6  *     Stefan Rompf <sux@loplof.de>
7  */
8 
9 #include <linux/module.h>
10 #include <linux/netdevice.h>
11 #include <linux/if.h>
12 #include <net/sock.h>
13 #include <net/pkt_sched.h>
14 #include <linux/rtnetlink.h>
15 #include <linux/jiffies.h>
16 #include <linux/spinlock.h>
17 #include <linux/workqueue.h>
18 #include <linux/bitops.h>
19 #include <linux/types.h>
20 
21 #include "dev.h"
22 
23 enum lw_bits {
24 	LW_URGENT = 0,
25 };
26 
27 static unsigned long linkwatch_flags;
28 static unsigned long linkwatch_nextevent;
29 
30 static void linkwatch_event(struct work_struct *dummy);
31 static DECLARE_DELAYED_WORK(linkwatch_work, linkwatch_event);
32 
33 static LIST_HEAD(lweventlist);
34 static DEFINE_SPINLOCK(lweventlist_lock);
35 
36 static unsigned int default_operstate(const struct net_device *dev)
37 {
38 	if (netif_testing(dev))
39 		return IF_OPER_TESTING;
40 
41 	/* Some uppers (DSA) have additional sources for being down, so
42 	 * first check whether lower is indeed the source of its down state.
43 	 */
44 	if (!netif_carrier_ok(dev)) {
45 		int iflink = dev_get_iflink(dev);
46 		struct net_device *peer;
47 
48 		if (iflink == dev->ifindex)
49 			return IF_OPER_DOWN;
50 
51 		peer = __dev_get_by_index(dev_net(dev), iflink);
52 		if (!peer)
53 			return IF_OPER_DOWN;
54 
55 		return netif_carrier_ok(peer) ? IF_OPER_DOWN :
56 						IF_OPER_LOWERLAYERDOWN;
57 	}
58 
59 	if (netif_dormant(dev))
60 		return IF_OPER_DORMANT;
61 
62 	return IF_OPER_UP;
63 }
64 
65 static void rfc2863_policy(struct net_device *dev)
66 {
67 	unsigned int operstate = default_operstate(dev);
68 
69 	if (operstate == READ_ONCE(dev->operstate))
70 		return;
71 
72 	switch(dev->link_mode) {
73 	case IF_LINK_MODE_TESTING:
74 		if (operstate == IF_OPER_UP)
75 			operstate = IF_OPER_TESTING;
76 		break;
77 
78 	case IF_LINK_MODE_DORMANT:
79 		if (operstate == IF_OPER_UP)
80 			operstate = IF_OPER_DORMANT;
81 		break;
82 	case IF_LINK_MODE_DEFAULT:
83 	default:
84 		break;
85 	}
86 
87 	WRITE_ONCE(dev->operstate, operstate);
88 }
89 
90 
91 void linkwatch_init_dev(struct net_device *dev)
92 {
93 	/* Handle pre-registration link state changes */
94 	if (!netif_carrier_ok(dev) || netif_dormant(dev) ||
95 	    netif_testing(dev))
96 		rfc2863_policy(dev);
97 }
98 
99 
100 static bool linkwatch_urgent_event(struct net_device *dev)
101 {
102 	if (!netif_running(dev))
103 		return false;
104 
105 	if (dev->ifindex != dev_get_iflink(dev))
106 		return true;
107 
108 	if (netif_is_lag_port(dev) || netif_is_lag_master(dev))
109 		return true;
110 
111 	return netif_carrier_ok(dev) &&	qdisc_tx_changing(dev);
112 }
113 
114 
115 static void linkwatch_add_event(struct net_device *dev)
116 {
117 	unsigned long flags;
118 
119 	spin_lock_irqsave(&lweventlist_lock, flags);
120 	if (list_empty(&dev->link_watch_list)) {
121 		list_add_tail(&dev->link_watch_list, &lweventlist);
122 		netdev_hold(dev, &dev->linkwatch_dev_tracker, GFP_ATOMIC);
123 	}
124 	spin_unlock_irqrestore(&lweventlist_lock, flags);
125 }
126 
127 
128 static void linkwatch_schedule_work(int urgent)
129 {
130 	unsigned long delay = linkwatch_nextevent - jiffies;
131 
132 	if (test_bit(LW_URGENT, &linkwatch_flags))
133 		return;
134 
135 	/* Minimise down-time: drop delay for up event. */
136 	if (urgent) {
137 		if (test_and_set_bit(LW_URGENT, &linkwatch_flags))
138 			return;
139 		delay = 0;
140 	}
141 
142 	/* If we wrap around we'll delay it by at most HZ. */
143 	if (delay > HZ)
144 		delay = 0;
145 
146 	/*
147 	 * If urgent, schedule immediate execution; otherwise, don't
148 	 * override the existing timer.
149 	 */
150 	if (test_bit(LW_URGENT, &linkwatch_flags))
151 		mod_delayed_work(system_wq, &linkwatch_work, 0);
152 	else
153 		schedule_delayed_work(&linkwatch_work, delay);
154 }
155 
156 
157 static void linkwatch_do_dev(struct net_device *dev)
158 {
159 	/*
160 	 * Make sure the above read is complete since it can be
161 	 * rewritten as soon as we clear the bit below.
162 	 */
163 	smp_mb__before_atomic();
164 
165 	/* We are about to handle this device,
166 	 * so new events can be accepted
167 	 */
168 	clear_bit(__LINK_STATE_LINKWATCH_PENDING, &dev->state);
169 
170 	rfc2863_policy(dev);
171 	if (dev->flags & IFF_UP) {
172 		if (netif_carrier_ok(dev))
173 			dev_activate(dev);
174 		else
175 			dev_deactivate(dev);
176 
177 		netdev_state_change(dev);
178 	}
179 	/* Note: our callers are responsible for calling netdev_tracker_free().
180 	 * This is the reason we use __dev_put() instead of dev_put().
181 	 */
182 	__dev_put(dev);
183 }
184 
185 static void __linkwatch_run_queue(int urgent_only)
186 {
187 #define MAX_DO_DEV_PER_LOOP	100
188 
189 	int do_dev = MAX_DO_DEV_PER_LOOP;
190 	/* Use a local list here since we add non-urgent
191 	 * events back to the global one when called with
192 	 * urgent_only=1.
193 	 */
194 	LIST_HEAD(wrk);
195 
196 	/* Give urgent case more budget */
197 	if (urgent_only)
198 		do_dev += MAX_DO_DEV_PER_LOOP;
199 
200 	/*
201 	 * Limit the number of linkwatch events to one
202 	 * per second so that a runaway driver does not
203 	 * cause a storm of messages on the netlink
204 	 * socket.  This limit does not apply to up events
205 	 * while the device qdisc is down.
206 	 */
207 	if (!urgent_only)
208 		linkwatch_nextevent = jiffies + HZ;
209 	/* Limit wrap-around effect on delay. */
210 	else if (time_after(linkwatch_nextevent, jiffies + HZ))
211 		linkwatch_nextevent = jiffies;
212 
213 	clear_bit(LW_URGENT, &linkwatch_flags);
214 
215 	spin_lock_irq(&lweventlist_lock);
216 	list_splice_init(&lweventlist, &wrk);
217 
218 	while (!list_empty(&wrk) && do_dev > 0) {
219 		struct net_device *dev;
220 
221 		dev = list_first_entry(&wrk, struct net_device, link_watch_list);
222 		list_del_init(&dev->link_watch_list);
223 
224 		if (!netif_device_present(dev) ||
225 		    (urgent_only && !linkwatch_urgent_event(dev))) {
226 			list_add_tail(&dev->link_watch_list, &lweventlist);
227 			continue;
228 		}
229 		/* We must free netdev tracker under
230 		 * the spinlock protection.
231 		 */
232 		netdev_tracker_free(dev, &dev->linkwatch_dev_tracker);
233 		spin_unlock_irq(&lweventlist_lock);
234 		linkwatch_do_dev(dev);
235 		do_dev--;
236 		spin_lock_irq(&lweventlist_lock);
237 	}
238 
239 	/* Add the remaining work back to lweventlist */
240 	list_splice_init(&wrk, &lweventlist);
241 
242 	if (!list_empty(&lweventlist))
243 		linkwatch_schedule_work(0);
244 	spin_unlock_irq(&lweventlist_lock);
245 }
246 
247 void linkwatch_sync_dev(struct net_device *dev)
248 {
249 	unsigned long flags;
250 	int clean = 0;
251 
252 	spin_lock_irqsave(&lweventlist_lock, flags);
253 	if (!list_empty(&dev->link_watch_list)) {
254 		list_del_init(&dev->link_watch_list);
255 		clean = 1;
256 		/* We must release netdev tracker under
257 		 * the spinlock protection.
258 		 */
259 		netdev_tracker_free(dev, &dev->linkwatch_dev_tracker);
260 	}
261 	spin_unlock_irqrestore(&lweventlist_lock, flags);
262 	if (clean)
263 		linkwatch_do_dev(dev);
264 }
265 
266 
267 /* Must be called with the rtnl semaphore held */
268 void linkwatch_run_queue(void)
269 {
270 	__linkwatch_run_queue(0);
271 }
272 
273 
274 static void linkwatch_event(struct work_struct *dummy)
275 {
276 	rtnl_lock();
277 	__linkwatch_run_queue(time_after(linkwatch_nextevent, jiffies));
278 	rtnl_unlock();
279 }
280 
281 
282 void linkwatch_fire_event(struct net_device *dev)
283 {
284 	bool urgent = linkwatch_urgent_event(dev);
285 
286 	if (!test_and_set_bit(__LINK_STATE_LINKWATCH_PENDING, &dev->state)) {
287 		linkwatch_add_event(dev);
288 	} else if (!urgent)
289 		return;
290 
291 	linkwatch_schedule_work(urgent);
292 }
293 EXPORT_SYMBOL(linkwatch_fire_event);
294