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 called from netdev_run_todo()/linkwatch_sync_dev(), 49 * dev_net(dev) can be already freed, and RTNL is not held. 50 */ 51 if (dev->reg_state == NETREG_UNREGISTERED || 52 iflink == dev->ifindex) 53 return IF_OPER_DOWN; 54 55 ASSERT_RTNL(); 56 peer = __dev_get_by_index(dev_net(dev), iflink); 57 if (!peer) 58 return IF_OPER_DOWN; 59 60 return netif_carrier_ok(peer) ? IF_OPER_DOWN : 61 IF_OPER_LOWERLAYERDOWN; 62 } 63 64 if (netif_dormant(dev)) 65 return IF_OPER_DORMANT; 66 67 return IF_OPER_UP; 68 } 69 70 71 static void rfc2863_policy(struct net_device *dev) 72 { 73 unsigned char operstate = default_operstate(dev); 74 75 if (operstate == READ_ONCE(dev->operstate)) 76 return; 77 78 write_lock(&dev_base_lock); 79 80 switch(dev->link_mode) { 81 case IF_LINK_MODE_TESTING: 82 if (operstate == IF_OPER_UP) 83 operstate = IF_OPER_TESTING; 84 break; 85 86 case IF_LINK_MODE_DORMANT: 87 if (operstate == IF_OPER_UP) 88 operstate = IF_OPER_DORMANT; 89 break; 90 case IF_LINK_MODE_DEFAULT: 91 default: 92 break; 93 } 94 95 WRITE_ONCE(dev->operstate, operstate); 96 97 write_unlock(&dev_base_lock); 98 } 99 100 101 void linkwatch_init_dev(struct net_device *dev) 102 { 103 /* Handle pre-registration link state changes */ 104 if (!netif_carrier_ok(dev) || netif_dormant(dev) || 105 netif_testing(dev)) 106 rfc2863_policy(dev); 107 } 108 109 110 static bool linkwatch_urgent_event(struct net_device *dev) 111 { 112 if (!netif_running(dev)) 113 return false; 114 115 if (dev->ifindex != dev_get_iflink(dev)) 116 return true; 117 118 if (netif_is_lag_port(dev) || netif_is_lag_master(dev)) 119 return true; 120 121 return netif_carrier_ok(dev) && qdisc_tx_changing(dev); 122 } 123 124 125 static void linkwatch_add_event(struct net_device *dev) 126 { 127 unsigned long flags; 128 129 spin_lock_irqsave(&lweventlist_lock, flags); 130 if (list_empty(&dev->link_watch_list)) { 131 list_add_tail(&dev->link_watch_list, &lweventlist); 132 netdev_hold(dev, &dev->linkwatch_dev_tracker, GFP_ATOMIC); 133 } 134 spin_unlock_irqrestore(&lweventlist_lock, flags); 135 } 136 137 138 static void linkwatch_schedule_work(int urgent) 139 { 140 unsigned long delay = linkwatch_nextevent - jiffies; 141 142 if (test_bit(LW_URGENT, &linkwatch_flags)) 143 return; 144 145 /* Minimise down-time: drop delay for up event. */ 146 if (urgent) { 147 if (test_and_set_bit(LW_URGENT, &linkwatch_flags)) 148 return; 149 delay = 0; 150 } 151 152 /* If we wrap around we'll delay it by at most HZ. */ 153 if (delay > HZ) 154 delay = 0; 155 156 /* 157 * If urgent, schedule immediate execution; otherwise, don't 158 * override the existing timer. 159 */ 160 if (test_bit(LW_URGENT, &linkwatch_flags)) 161 mod_delayed_work(system_unbound_wq, &linkwatch_work, 0); 162 else 163 queue_delayed_work(system_unbound_wq, &linkwatch_work, delay); 164 } 165 166 167 static void linkwatch_do_dev(struct net_device *dev) 168 { 169 /* 170 * Make sure the above read is complete since it can be 171 * rewritten as soon as we clear the bit below. 172 */ 173 smp_mb__before_atomic(); 174 175 /* We are about to handle this device, 176 * so new events can be accepted 177 */ 178 clear_bit(__LINK_STATE_LINKWATCH_PENDING, &dev->state); 179 180 rfc2863_policy(dev); 181 if (dev->flags & IFF_UP) { 182 if (netif_carrier_ok(dev)) 183 dev_activate(dev); 184 else 185 dev_deactivate(dev); 186 187 netdev_state_change(dev); 188 } 189 /* Note: our callers are responsible for calling netdev_tracker_free(). 190 * This is the reason we use __dev_put() instead of dev_put(). 191 */ 192 __dev_put(dev); 193 } 194 195 static void __linkwatch_run_queue(int urgent_only) 196 { 197 #define MAX_DO_DEV_PER_LOOP 100 198 199 int do_dev = MAX_DO_DEV_PER_LOOP; 200 struct net_device *dev; 201 LIST_HEAD(wrk); 202 203 /* Give urgent case more budget */ 204 if (urgent_only) 205 do_dev += MAX_DO_DEV_PER_LOOP; 206 207 /* 208 * Limit the number of linkwatch events to one 209 * per second so that a runaway driver does not 210 * cause a storm of messages on the netlink 211 * socket. This limit does not apply to up events 212 * while the device qdisc is down. 213 */ 214 if (!urgent_only) 215 linkwatch_nextevent = jiffies + HZ; 216 /* Limit wrap-around effect on delay. */ 217 else if (time_after(linkwatch_nextevent, jiffies + HZ)) 218 linkwatch_nextevent = jiffies; 219 220 clear_bit(LW_URGENT, &linkwatch_flags); 221 222 spin_lock_irq(&lweventlist_lock); 223 list_splice_init(&lweventlist, &wrk); 224 225 while (!list_empty(&wrk) && do_dev > 0) { 226 227 dev = list_first_entry(&wrk, struct net_device, link_watch_list); 228 list_del_init(&dev->link_watch_list); 229 230 if (!netif_device_present(dev) || 231 (urgent_only && !linkwatch_urgent_event(dev))) { 232 list_add_tail(&dev->link_watch_list, &lweventlist); 233 continue; 234 } 235 /* We must free netdev tracker under 236 * the spinlock protection. 237 */ 238 netdev_tracker_free(dev, &dev->linkwatch_dev_tracker); 239 spin_unlock_irq(&lweventlist_lock); 240 linkwatch_do_dev(dev); 241 do_dev--; 242 spin_lock_irq(&lweventlist_lock); 243 } 244 245 /* Add the remaining work back to lweventlist */ 246 list_splice_init(&wrk, &lweventlist); 247 248 if (!list_empty(&lweventlist)) 249 linkwatch_schedule_work(0); 250 spin_unlock_irq(&lweventlist_lock); 251 } 252 253 void linkwatch_forget_dev(struct net_device *dev) 254 { 255 unsigned long flags; 256 int clean = 0; 257 258 spin_lock_irqsave(&lweventlist_lock, flags); 259 if (!list_empty(&dev->link_watch_list)) { 260 list_del_init(&dev->link_watch_list); 261 clean = 1; 262 /* We must release netdev tracker under 263 * the spinlock protection. 264 */ 265 netdev_tracker_free(dev, &dev->linkwatch_dev_tracker); 266 } 267 spin_unlock_irqrestore(&lweventlist_lock, flags); 268 if (clean) 269 linkwatch_do_dev(dev); 270 } 271 272 273 /* Must be called with the rtnl semaphore held */ 274 void linkwatch_run_queue(void) 275 { 276 __linkwatch_run_queue(0); 277 } 278 279 280 static void linkwatch_event(struct work_struct *dummy) 281 { 282 rtnl_lock(); 283 __linkwatch_run_queue(time_after(linkwatch_nextevent, jiffies)); 284 rtnl_unlock(); 285 } 286 287 288 void linkwatch_fire_event(struct net_device *dev) 289 { 290 bool urgent = linkwatch_urgent_event(dev); 291 292 if (!test_and_set_bit(__LINK_STATE_LINKWATCH_PENDING, &dev->state)) { 293 linkwatch_add_event(dev); 294 } else if (!urgent) 295 return; 296 297 linkwatch_schedule_work(urgent); 298 } 299 EXPORT_SYMBOL(linkwatch_fire_event); 300