xref: /openbmc/linux/net/core/link_watch.c (revision 5cd17f0e)
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 char 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 
66 static void rfc2863_policy(struct net_device *dev)
67 {
68 	unsigned char operstate = default_operstate(dev);
69 
70 	if (operstate == READ_ONCE(dev->operstate))
71 		return;
72 
73 	write_lock(&dev_base_lock);
74 
75 	switch(dev->link_mode) {
76 	case IF_LINK_MODE_TESTING:
77 		if (operstate == IF_OPER_UP)
78 			operstate = IF_OPER_TESTING;
79 		break;
80 
81 	case IF_LINK_MODE_DORMANT:
82 		if (operstate == IF_OPER_UP)
83 			operstate = IF_OPER_DORMANT;
84 		break;
85 	case IF_LINK_MODE_DEFAULT:
86 	default:
87 		break;
88 	}
89 
90 	WRITE_ONCE(dev->operstate, operstate);
91 
92 	write_unlock(&dev_base_lock);
93 }
94 
95 
96 void linkwatch_init_dev(struct net_device *dev)
97 {
98 	/* Handle pre-registration link state changes */
99 	if (!netif_carrier_ok(dev) || netif_dormant(dev) ||
100 	    netif_testing(dev))
101 		rfc2863_policy(dev);
102 }
103 
104 
105 static bool linkwatch_urgent_event(struct net_device *dev)
106 {
107 	if (!netif_running(dev))
108 		return false;
109 
110 	if (dev->ifindex != dev_get_iflink(dev))
111 		return true;
112 
113 	if (netif_is_lag_port(dev) || netif_is_lag_master(dev))
114 		return true;
115 
116 	return netif_carrier_ok(dev) &&	qdisc_tx_changing(dev);
117 }
118 
119 
120 static void linkwatch_add_event(struct net_device *dev)
121 {
122 	unsigned long flags;
123 
124 	spin_lock_irqsave(&lweventlist_lock, flags);
125 	if (list_empty(&dev->link_watch_list)) {
126 		list_add_tail(&dev->link_watch_list, &lweventlist);
127 		netdev_hold(dev, &dev->linkwatch_dev_tracker, GFP_ATOMIC);
128 	}
129 	spin_unlock_irqrestore(&lweventlist_lock, flags);
130 }
131 
132 
133 static void linkwatch_schedule_work(int urgent)
134 {
135 	unsigned long delay = linkwatch_nextevent - jiffies;
136 
137 	if (test_bit(LW_URGENT, &linkwatch_flags))
138 		return;
139 
140 	/* Minimise down-time: drop delay for up event. */
141 	if (urgent) {
142 		if (test_and_set_bit(LW_URGENT, &linkwatch_flags))
143 			return;
144 		delay = 0;
145 	}
146 
147 	/* If we wrap around we'll delay it by at most HZ. */
148 	if (delay > HZ)
149 		delay = 0;
150 
151 	/*
152 	 * If urgent, schedule immediate execution; otherwise, don't
153 	 * override the existing timer.
154 	 */
155 	if (test_bit(LW_URGENT, &linkwatch_flags))
156 		mod_delayed_work(system_wq, &linkwatch_work, 0);
157 	else
158 		schedule_delayed_work(&linkwatch_work, delay);
159 }
160 
161 
162 static void linkwatch_do_dev(struct net_device *dev)
163 {
164 	/*
165 	 * Make sure the above read is complete since it can be
166 	 * rewritten as soon as we clear the bit below.
167 	 */
168 	smp_mb__before_atomic();
169 
170 	/* We are about to handle this device,
171 	 * so new events can be accepted
172 	 */
173 	clear_bit(__LINK_STATE_LINKWATCH_PENDING, &dev->state);
174 
175 	rfc2863_policy(dev);
176 	if (dev->flags & IFF_UP) {
177 		if (netif_carrier_ok(dev))
178 			dev_activate(dev);
179 		else
180 			dev_deactivate(dev);
181 
182 		netdev_state_change(dev);
183 	}
184 	/* Note: our callers are responsible for calling netdev_tracker_free().
185 	 * This is the reason we use __dev_put() instead of dev_put().
186 	 */
187 	__dev_put(dev);
188 }
189 
190 static void __linkwatch_run_queue(int urgent_only)
191 {
192 #define MAX_DO_DEV_PER_LOOP	100
193 
194 	int do_dev = MAX_DO_DEV_PER_LOOP;
195 	struct net_device *dev;
196 	LIST_HEAD(wrk);
197 
198 	/* Give urgent case more budget */
199 	if (urgent_only)
200 		do_dev += MAX_DO_DEV_PER_LOOP;
201 
202 	/*
203 	 * Limit the number of linkwatch events to one
204 	 * per second so that a runaway driver does not
205 	 * cause a storm of messages on the netlink
206 	 * socket.  This limit does not apply to up events
207 	 * while the device qdisc is down.
208 	 */
209 	if (!urgent_only)
210 		linkwatch_nextevent = jiffies + HZ;
211 	/* Limit wrap-around effect on delay. */
212 	else if (time_after(linkwatch_nextevent, jiffies + HZ))
213 		linkwatch_nextevent = jiffies;
214 
215 	clear_bit(LW_URGENT, &linkwatch_flags);
216 
217 	spin_lock_irq(&lweventlist_lock);
218 	list_splice_init(&lweventlist, &wrk);
219 
220 	while (!list_empty(&wrk) && do_dev > 0) {
221 
222 		dev = list_first_entry(&wrk, struct net_device, link_watch_list);
223 		list_del_init(&dev->link_watch_list);
224 
225 		if (!netif_device_present(dev) ||
226 		    (urgent_only && !linkwatch_urgent_event(dev))) {
227 			list_add_tail(&dev->link_watch_list, &lweventlist);
228 			continue;
229 		}
230 		/* We must free netdev tracker under
231 		 * the spinlock protection.
232 		 */
233 		netdev_tracker_free(dev, &dev->linkwatch_dev_tracker);
234 		spin_unlock_irq(&lweventlist_lock);
235 		linkwatch_do_dev(dev);
236 		do_dev--;
237 		spin_lock_irq(&lweventlist_lock);
238 	}
239 
240 	/* Add the remaining work back to lweventlist */
241 	list_splice_init(&wrk, &lweventlist);
242 
243 	if (!list_empty(&lweventlist))
244 		linkwatch_schedule_work(0);
245 	spin_unlock_irq(&lweventlist_lock);
246 }
247 
248 void linkwatch_forget_dev(struct net_device *dev)
249 {
250 	unsigned long flags;
251 	int clean = 0;
252 
253 	spin_lock_irqsave(&lweventlist_lock, flags);
254 	if (!list_empty(&dev->link_watch_list)) {
255 		list_del_init(&dev->link_watch_list);
256 		clean = 1;
257 		/* We must release netdev tracker under
258 		 * the spinlock protection.
259 		 */
260 		netdev_tracker_free(dev, &dev->linkwatch_dev_tracker);
261 	}
262 	spin_unlock_irqrestore(&lweventlist_lock, flags);
263 	if (clean)
264 		linkwatch_do_dev(dev);
265 }
266 
267 
268 /* Must be called with the rtnl semaphore held */
269 void linkwatch_run_queue(void)
270 {
271 	__linkwatch_run_queue(0);
272 }
273 
274 
275 static void linkwatch_event(struct work_struct *dummy)
276 {
277 	rtnl_lock();
278 	__linkwatch_run_queue(time_after(linkwatch_nextevent, jiffies));
279 	rtnl_unlock();
280 }
281 
282 
283 void linkwatch_fire_event(struct net_device *dev)
284 {
285 	bool urgent = linkwatch_urgent_event(dev);
286 
287 	if (!test_and_set_bit(__LINK_STATE_LINKWATCH_PENDING, &dev->state)) {
288 		linkwatch_add_event(dev);
289 	} else if (!urgent)
290 		return;
291 
292 	linkwatch_schedule_work(urgent);
293 }
294 EXPORT_SYMBOL(linkwatch_fire_event);
295