Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * net/core/dst.c Protocol independent destination cache. |
| 3 | * |
| 4 | * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru> |
| 5 | * |
| 6 | */ |
| 7 | |
| 8 | #include <linux/bitops.h> |
| 9 | #include <linux/errno.h> |
| 10 | #include <linux/init.h> |
| 11 | #include <linux/kernel.h> |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 12 | #include <linux/workqueue.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 13 | #include <linux/mm.h> |
| 14 | #include <linux/module.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 15 | #include <linux/slab.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 16 | #include <linux/netdevice.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 17 | #include <linux/skbuff.h> |
| 18 | #include <linux/string.h> |
| 19 | #include <linux/types.h> |
Eric W. Biederman | e9dc865 | 2007-09-12 13:02:17 +0200 | [diff] [blame] | 20 | #include <net/net_namespace.h> |
Eric Dumazet | 2fc1b5d | 2010-02-08 15:00:39 -0800 | [diff] [blame] | 21 | #include <linux/sched.h> |
Linus Torvalds | 268bb0c | 2011-05-20 12:50:29 -0700 | [diff] [blame] | 22 | #include <linux/prefetch.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 23 | |
| 24 | #include <net/dst.h> |
| 25 | |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 26 | /* |
| 27 | * Theory of operations: |
| 28 | * 1) We use a list, protected by a spinlock, to add |
| 29 | * new entries from both BH and non-BH context. |
| 30 | * 2) In order to keep spinlock held for a small delay, |
| 31 | * we use a second list where are stored long lived |
| 32 | * entries, that are handled by the garbage collect thread |
| 33 | * fired by a workqueue. |
| 34 | * 3) This list is guarded by a mutex, |
| 35 | * so that the gc_task and dst_dev_event() can be synchronized. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 36 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 37 | |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 38 | /* |
| 39 | * We want to keep lock & list close together |
| 40 | * to dirty as few cache lines as possible in __dst_free(). |
| 41 | * As this is not a very strong hint, we dont force an alignment on SMP. |
| 42 | */ |
| 43 | static struct { |
| 44 | spinlock_t lock; |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 45 | struct dst_entry *list; |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 46 | unsigned long timer_inc; |
| 47 | unsigned long timer_expires; |
| 48 | } dst_garbage = { |
| 49 | .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock), |
| 50 | .timer_inc = DST_GC_MAX, |
| 51 | }; |
| 52 | static void dst_gc_task(struct work_struct *work); |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 53 | static void ___dst_free(struct dst_entry *dst); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 54 | |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 55 | static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 56 | |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 57 | static DEFINE_MUTEX(dst_gc_mutex); |
| 58 | /* |
| 59 | * long lived entries are maintained in this list, guarded by dst_gc_mutex |
| 60 | */ |
| 61 | static struct dst_entry *dst_busy_list; |
| 62 | |
| 63 | static void dst_gc_task(struct work_struct *work) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 64 | { |
| 65 | int delayed = 0; |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 66 | int work_performed = 0; |
| 67 | unsigned long expires = ~0L; |
| 68 | struct dst_entry *dst, *next, head; |
| 69 | struct dst_entry *last = &head; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 70 | |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 71 | mutex_lock(&dst_gc_mutex); |
| 72 | next = dst_busy_list; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 73 | |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 74 | loop: |
| 75 | while ((dst = next) != NULL) { |
| 76 | next = dst->next; |
| 77 | prefetch(&next->next); |
Eric Dumazet | 2fc1b5d | 2010-02-08 15:00:39 -0800 | [diff] [blame] | 78 | cond_resched(); |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 79 | if (likely(atomic_read(&dst->__refcnt))) { |
| 80 | last->next = dst; |
| 81 | last = dst; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 82 | delayed++; |
| 83 | continue; |
| 84 | } |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 85 | work_performed++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 86 | |
| 87 | dst = dst_destroy(dst); |
| 88 | if (dst) { |
| 89 | /* NOHASH and still referenced. Unless it is already |
| 90 | * on gc list, invalidate it and add to gc list. |
| 91 | * |
| 92 | * Note: this is temporary. Actually, NOHASH dst's |
| 93 | * must be obsoleted when parent is obsoleted. |
| 94 | * But we do not have state "obsoleted, but |
| 95 | * referenced by parent", so it is right. |
| 96 | */ |
David S. Miller | f5b0a87 | 2012-07-19 12:31:33 -0700 | [diff] [blame] | 97 | if (dst->obsolete > 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 98 | continue; |
| 99 | |
| 100 | ___dst_free(dst); |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 101 | dst->next = next; |
| 102 | next = dst; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 103 | } |
| 104 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 105 | |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 106 | spin_lock_bh(&dst_garbage.lock); |
| 107 | next = dst_garbage.list; |
| 108 | if (next) { |
| 109 | dst_garbage.list = NULL; |
| 110 | spin_unlock_bh(&dst_garbage.lock); |
| 111 | goto loop; |
| 112 | } |
| 113 | last->next = NULL; |
| 114 | dst_busy_list = head.next; |
| 115 | if (!dst_busy_list) |
| 116 | dst_garbage.timer_inc = DST_GC_MAX; |
| 117 | else { |
| 118 | /* |
| 119 | * if we freed less than 1/10 of delayed entries, |
| 120 | * we can sleep longer. |
| 121 | */ |
| 122 | if (work_performed <= delayed/10) { |
| 123 | dst_garbage.timer_expires += dst_garbage.timer_inc; |
| 124 | if (dst_garbage.timer_expires > DST_GC_MAX) |
| 125 | dst_garbage.timer_expires = DST_GC_MAX; |
| 126 | dst_garbage.timer_inc += DST_GC_INC; |
| 127 | } else { |
| 128 | dst_garbage.timer_inc = DST_GC_INC; |
| 129 | dst_garbage.timer_expires = DST_GC_MIN; |
| 130 | } |
| 131 | expires = dst_garbage.timer_expires; |
| 132 | /* |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 133 | * if the next desired timer is more than 4 seconds in the |
| 134 | * future then round the timer to whole seconds |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 135 | */ |
| 136 | if (expires > 4*HZ) |
| 137 | expires = round_jiffies_relative(expires); |
| 138 | schedule_delayed_work(&dst_gc_work, expires); |
| 139 | } |
| 140 | |
| 141 | spin_unlock_bh(&dst_garbage.lock); |
| 142 | mutex_unlock(&dst_gc_mutex); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 143 | } |
| 144 | |
Herbert Xu | 352e512 | 2007-11-13 21:34:06 -0800 | [diff] [blame] | 145 | int dst_discard(struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 146 | { |
| 147 | kfree_skb(skb); |
| 148 | return 0; |
| 149 | } |
Herbert Xu | 352e512 | 2007-11-13 21:34:06 -0800 | [diff] [blame] | 150 | EXPORT_SYMBOL(dst_discard); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 151 | |
Eric Dumazet | a37e6e3 | 2012-08-07 10:55:45 +0000 | [diff] [blame] | 152 | const u32 dst_default_metrics[RTAX_MAX + 1] = { |
| 153 | /* This initializer is needed to force linker to place this variable |
| 154 | * into const section. Otherwise it might end into bss section. |
| 155 | * We really want to avoid false sharing on this variable, and catch |
| 156 | * any writes on it. |
| 157 | */ |
| 158 | [RTAX_MAX] = 0xdeadbeef, |
| 159 | }; |
| 160 | |
David S. Miller | 62fa8a8 | 2011-01-26 20:51:05 -0800 | [diff] [blame] | 161 | |
David S. Miller | 5c1e6aa | 2011-04-28 14:13:38 -0700 | [diff] [blame] | 162 | void *dst_alloc(struct dst_ops *ops, struct net_device *dev, |
David S. Miller | 5110effe | 2012-07-02 02:21:03 -0700 | [diff] [blame] | 163 | int initial_ref, int initial_obsolete, unsigned short flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 164 | { |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 165 | struct dst_entry *dst; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 166 | |
Eric Dumazet | fc66f95 | 2010-10-08 06:37:34 +0000 | [diff] [blame] | 167 | if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) { |
Daniel Lezcano | 569d364 | 2008-01-18 03:56:57 -0800 | [diff] [blame] | 168 | if (ops->gc(ops)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 169 | return NULL; |
| 170 | } |
David S. Miller | cf91166 | 2011-04-28 14:31:47 -0700 | [diff] [blame] | 171 | dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 172 | if (!dst) |
| 173 | return NULL; |
David S. Miller | cf91166 | 2011-04-28 14:31:47 -0700 | [diff] [blame] | 174 | dst->child = NULL; |
David S. Miller | 5c1e6aa | 2011-04-28 14:13:38 -0700 | [diff] [blame] | 175 | dst->dev = dev; |
| 176 | if (dev) |
| 177 | dev_hold(dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 178 | dst->ops = ops; |
David S. Miller | 62fa8a8 | 2011-01-26 20:51:05 -0800 | [diff] [blame] | 179 | dst_init_metrics(dst, dst_default_metrics, true); |
David S. Miller | cf91166 | 2011-04-28 14:31:47 -0700 | [diff] [blame] | 180 | dst->expires = 0UL; |
David S. Miller | 5c1e6aa | 2011-04-28 14:13:38 -0700 | [diff] [blame] | 181 | dst->path = dst; |
David S. Miller | cf91166 | 2011-04-28 14:31:47 -0700 | [diff] [blame] | 182 | #ifdef CONFIG_XFRM |
| 183 | dst->xfrm = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 184 | #endif |
David S. Miller | 5c1e6aa | 2011-04-28 14:13:38 -0700 | [diff] [blame] | 185 | dst->input = dst_discard; |
| 186 | dst->output = dst_discard; |
David S. Miller | cf91166 | 2011-04-28 14:31:47 -0700 | [diff] [blame] | 187 | dst->error = 0; |
David S. Miller | 5c1e6aa | 2011-04-28 14:13:38 -0700 | [diff] [blame] | 188 | dst->obsolete = initial_obsolete; |
David S. Miller | cf91166 | 2011-04-28 14:31:47 -0700 | [diff] [blame] | 189 | dst->header_len = 0; |
| 190 | dst->trailer_len = 0; |
| 191 | #ifdef CONFIG_IP_ROUTE_CLASSID |
| 192 | dst->tclassid = 0; |
| 193 | #endif |
David S. Miller | 5c1e6aa | 2011-04-28 14:13:38 -0700 | [diff] [blame] | 194 | atomic_set(&dst->__refcnt, initial_ref); |
David S. Miller | cf91166 | 2011-04-28 14:31:47 -0700 | [diff] [blame] | 195 | dst->__use = 0; |
David S. Miller | 5c1e6aa | 2011-04-28 14:13:38 -0700 | [diff] [blame] | 196 | dst->lastuse = jiffies; |
| 197 | dst->flags = flags; |
David S. Miller | 5110effe | 2012-07-02 02:21:03 -0700 | [diff] [blame] | 198 | dst->pending_confirm = 0; |
David S. Miller | cf91166 | 2011-04-28 14:31:47 -0700 | [diff] [blame] | 199 | dst->next = NULL; |
David S. Miller | 957c665 | 2011-06-24 15:25:00 -0700 | [diff] [blame] | 200 | if (!(flags & DST_NOCOUNT)) |
| 201 | dst_entries_add(ops, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 202 | return dst; |
| 203 | } |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 204 | EXPORT_SYMBOL(dst_alloc); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 205 | |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 206 | static void ___dst_free(struct dst_entry *dst) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 207 | { |
| 208 | /* The first case (dev==NULL) is required, when |
| 209 | protocol module is unloaded. |
| 210 | */ |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 211 | if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) |
Denis Cheng | c4b1010 | 2007-06-05 00:06:57 -0700 | [diff] [blame] | 212 | dst->input = dst->output = dst_discard; |
David S. Miller | f5b0a87 | 2012-07-19 12:31:33 -0700 | [diff] [blame] | 213 | dst->obsolete = DST_OBSOLETE_DEAD; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 214 | } |
| 215 | |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 216 | void __dst_free(struct dst_entry *dst) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 217 | { |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 218 | spin_lock_bh(&dst_garbage.lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 219 | ___dst_free(dst); |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 220 | dst->next = dst_garbage.list; |
| 221 | dst_garbage.list = dst; |
| 222 | if (dst_garbage.timer_inc > DST_GC_INC) { |
| 223 | dst_garbage.timer_inc = DST_GC_INC; |
| 224 | dst_garbage.timer_expires = DST_GC_MIN; |
Tejun Heo | 41f63c5 | 2012-08-03 10:30:47 -0700 | [diff] [blame] | 225 | mod_delayed_work(system_wq, &dst_gc_work, |
| 226 | dst_garbage.timer_expires); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 227 | } |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 228 | spin_unlock_bh(&dst_garbage.lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 229 | } |
Nicolas Dichtel | d79d991 | 2010-07-19 23:51:38 +0000 | [diff] [blame] | 230 | EXPORT_SYMBOL(__dst_free); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 231 | |
| 232 | struct dst_entry *dst_destroy(struct dst_entry * dst) |
| 233 | { |
| 234 | struct dst_entry *child; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 235 | |
| 236 | smp_rmb(); |
| 237 | |
| 238 | again: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 239 | child = dst->child; |
| 240 | |
David S. Miller | 957c665 | 2011-06-24 15:25:00 -0700 | [diff] [blame] | 241 | if (!(dst->flags & DST_NOCOUNT)) |
| 242 | dst_entries_add(dst->ops, -1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 243 | |
| 244 | if (dst->ops->destroy) |
| 245 | dst->ops->destroy(dst); |
| 246 | if (dst->dev) |
| 247 | dev_put(dst->dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 248 | kmem_cache_free(dst->ops->kmem_cachep, dst); |
| 249 | |
| 250 | dst = child; |
| 251 | if (dst) { |
Herbert Xu | 6775cab | 2005-04-16 15:24:10 -0700 | [diff] [blame] | 252 | int nohash = dst->flags & DST_NOHASH; |
| 253 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 254 | if (atomic_dec_and_test(&dst->__refcnt)) { |
| 255 | /* We were real parent of this dst, so kill child. */ |
Herbert Xu | 6775cab | 2005-04-16 15:24:10 -0700 | [diff] [blame] | 256 | if (nohash) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 257 | goto again; |
| 258 | } else { |
| 259 | /* Child is still referenced, return it for freeing. */ |
Herbert Xu | 6775cab | 2005-04-16 15:24:10 -0700 | [diff] [blame] | 260 | if (nohash) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 261 | return dst; |
| 262 | /* Child is still in his hash table */ |
| 263 | } |
| 264 | } |
| 265 | return NULL; |
| 266 | } |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 267 | EXPORT_SYMBOL(dst_destroy); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 268 | |
Ilpo Järvinen | 8d33086 | 2008-03-27 17:53:31 -0700 | [diff] [blame] | 269 | void dst_release(struct dst_entry *dst) |
| 270 | { |
| 271 | if (dst) { |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 272 | int newrefcnt; |
Eric Dumazet | ef711cf | 2008-11-14 00:53:54 -0800 | [diff] [blame] | 273 | |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 274 | newrefcnt = atomic_dec_return(&dst->__refcnt); |
| 275 | WARN_ON(newrefcnt < 0); |
Eric Dumazet | 27b75c9 | 2010-10-15 05:44:11 +0000 | [diff] [blame] | 276 | if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) { |
| 277 | dst = dst_destroy(dst); |
| 278 | if (dst) |
| 279 | __dst_free(dst); |
| 280 | } |
Ilpo Järvinen | 8d33086 | 2008-03-27 17:53:31 -0700 | [diff] [blame] | 281 | } |
| 282 | } |
| 283 | EXPORT_SYMBOL(dst_release); |
| 284 | |
David S. Miller | 62fa8a8 | 2011-01-26 20:51:05 -0800 | [diff] [blame] | 285 | u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old) |
| 286 | { |
| 287 | u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC); |
| 288 | |
| 289 | if (p) { |
| 290 | u32 *old_p = __DST_METRICS_PTR(old); |
| 291 | unsigned long prev, new; |
| 292 | |
| 293 | memcpy(p, old_p, sizeof(u32) * RTAX_MAX); |
| 294 | |
| 295 | new = (unsigned long) p; |
| 296 | prev = cmpxchg(&dst->_metrics, old, new); |
| 297 | |
| 298 | if (prev != old) { |
| 299 | kfree(p); |
| 300 | p = __DST_METRICS_PTR(prev); |
| 301 | if (prev & DST_METRICS_READ_ONLY) |
| 302 | p = NULL; |
| 303 | } |
| 304 | } |
| 305 | return p; |
| 306 | } |
| 307 | EXPORT_SYMBOL(dst_cow_metrics_generic); |
| 308 | |
| 309 | /* Caller asserts that dst_metrics_read_only(dst) is false. */ |
| 310 | void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old) |
| 311 | { |
| 312 | unsigned long prev, new; |
| 313 | |
Eric Dumazet | b30c516 | 2011-05-24 13:29:50 -0400 | [diff] [blame] | 314 | new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY; |
David S. Miller | 62fa8a8 | 2011-01-26 20:51:05 -0800 | [diff] [blame] | 315 | prev = cmpxchg(&dst->_metrics, old, new); |
| 316 | if (prev == old) |
| 317 | kfree(__DST_METRICS_PTR(old)); |
| 318 | } |
| 319 | EXPORT_SYMBOL(__dst_destroy_metrics_generic); |
| 320 | |
Eric Dumazet | 27b75c9 | 2010-10-15 05:44:11 +0000 | [diff] [blame] | 321 | /** |
| 322 | * skb_dst_set_noref - sets skb dst, without a reference |
| 323 | * @skb: buffer |
| 324 | * @dst: dst entry |
| 325 | * |
| 326 | * Sets skb dst, assuming a reference was not taken on dst |
| 327 | * skb_dst_drop() should not dst_release() this dst |
| 328 | */ |
| 329 | void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst) |
| 330 | { |
| 331 | WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); |
| 332 | /* If dst not in cache, we must take a reference, because |
| 333 | * dst_release() will destroy dst as soon as its refcount becomes zero |
| 334 | */ |
| 335 | if (unlikely(dst->flags & DST_NOCACHE)) { |
| 336 | dst_hold(dst); |
| 337 | skb_dst_set(skb, dst); |
| 338 | } else { |
| 339 | skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF; |
| 340 | } |
| 341 | } |
| 342 | EXPORT_SYMBOL(skb_dst_set_noref); |
| 343 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 344 | /* Dirty hack. We did it in 2.2 (in __dst_free), |
| 345 | * we have _very_ good reasons not to repeat |
| 346 | * this mistake in 2.3, but we have no choice |
| 347 | * now. _It_ _is_ _explicit_ _deliberate_ |
| 348 | * _race_ _condition_. |
| 349 | * |
| 350 | * Commented and originally written by Alexey. |
| 351 | */ |
stephen hemminger | 5611551 | 2010-04-12 07:38:05 +0000 | [diff] [blame] | 352 | static void dst_ifdown(struct dst_entry *dst, struct net_device *dev, |
| 353 | int unregister) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 354 | { |
| 355 | if (dst->ops->ifdown) |
| 356 | dst->ops->ifdown(dst, dev, unregister); |
| 357 | |
| 358 | if (dev != dst->dev) |
| 359 | return; |
| 360 | |
| 361 | if (!unregister) { |
Denis Cheng | c4b1010 | 2007-06-05 00:06:57 -0700 | [diff] [blame] | 362 | dst->input = dst->output = dst_discard; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 363 | } else { |
YOSHIFUJI Hideaki | c346dca | 2008-03-25 21:47:49 +0900 | [diff] [blame] | 364 | dst->dev = dev_net(dst->dev)->loopback_dev; |
Daniel Lezcano | de3cb74 | 2007-09-25 19:16:28 -0700 | [diff] [blame] | 365 | dev_hold(dst->dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 366 | dev_put(dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 367 | } |
| 368 | } |
| 369 | |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 370 | static int dst_dev_event(struct notifier_block *this, unsigned long event, |
| 371 | void *ptr) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 372 | { |
| 373 | struct net_device *dev = ptr; |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 374 | struct dst_entry *dst, *last = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 375 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 376 | switch (event) { |
Eric Dumazet | 0115e8e | 2012-08-22 17:19:46 +0000 | [diff] [blame] | 377 | case NETDEV_UNREGISTER_FINAL: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 378 | case NETDEV_DOWN: |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 379 | mutex_lock(&dst_gc_mutex); |
| 380 | for (dst = dst_busy_list; dst; dst = dst->next) { |
| 381 | last = dst; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 382 | dst_ifdown(dst, dev, event != NETDEV_DOWN); |
| 383 | } |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 384 | |
| 385 | spin_lock_bh(&dst_garbage.lock); |
| 386 | dst = dst_garbage.list; |
| 387 | dst_garbage.list = NULL; |
| 388 | spin_unlock_bh(&dst_garbage.lock); |
| 389 | |
| 390 | if (last) |
| 391 | last->next = dst; |
| 392 | else |
| 393 | dst_busy_list = dst; |
laurent chavey | 598ed93 | 2010-03-29 10:41:36 +0000 | [diff] [blame] | 394 | for (; dst; dst = dst->next) |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 395 | dst_ifdown(dst, dev, event != NETDEV_DOWN); |
Eric Dumazet | 86bba26 | 2007-09-12 14:29:01 +0200 | [diff] [blame] | 396 | mutex_unlock(&dst_gc_mutex); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 397 | break; |
| 398 | } |
| 399 | return NOTIFY_DONE; |
| 400 | } |
| 401 | |
| 402 | static struct notifier_block dst_dev_notifier = { |
| 403 | .notifier_call = dst_dev_event, |
Eric Dumazet | 332dd96 | 2010-11-09 11:46:33 -0800 | [diff] [blame] | 404 | .priority = -10, /* must be called after other network notifiers */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 405 | }; |
| 406 | |
| 407 | void __init dst_init(void) |
| 408 | { |
| 409 | register_netdevice_notifier(&dst_dev_notifier); |
| 410 | } |