net/core/dst.c

   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>
  12 #include <linux/workqueue.h>
  13 #include <linux/mm.h>
  14 #include <linux/module.h>
  15 #include <linux/slab.h>
  16 #include <linux/netdevice.h>
  17 #include <linux/skbuff.h>
  18 #include <linux/string.h>
  19 #include <linux/types.h>
  20 #include <net/net_namespace.h>
  21 #include <linux/sched.h>
  22
  23 #include <net/dst.h>
  24
  25 /*
  26  * Theory of operations:
  27  * 1) We use a list, protected by a spinlock, to add
  28  *    new entries from both BH and non-BH context.
  29  * 2) In order to keep spinlock held for a small delay,
  30  *    we use a second list where are stored long lived
  31  *    entries, that are handled by the garbage collect thread
  32  *    fired by a workqueue.
  33  * 3) This list is guarded by a mutex,
  34  *    so that the gc_task and dst_dev_event() can be synchronized.
  35  */
  36 #if RT_CACHE_DEBUG >= 2
  37 static atomic_t                  dst_total = ATOMIC_INIT(0);
  38 #endif
  39
  40 /*
  41  * We want to keep lock & list close together
  42  * to dirty as few cache lines as possible in __dst_free().
  43  * As this is not a very strong hint, we dont force an alignment on SMP.
  44  */
  45 static struct {
  46         spinlock_t              lock;
  47         struct dst_entry        *list;
  48         unsigned long           timer_inc;
  49         unsigned long           timer_expires;
  50 } dst_garbage = {
  51         .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock),
  52         .timer_inc = DST_GC_MAX,
  53 };
  54 static void dst_gc_task(struct work_struct *work);
  55 static void ___dst_free(struct dst_entry *dst);
  56
  57 static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
  58
  59 static DEFINE_MUTEX(dst_gc_mutex);
  60 /*
  61  * long lived entries are maintained in this list, guarded by dst_gc_mutex
  62  */
  63 static struct dst_entry         *dst_busy_list;
  64
  65 static void dst_gc_task(struct work_struct *work)
  66 {
  67         int    delayed = 0;
  68         int    work_performed = 0;
  69         unsigned long expires = ~0L;
  70         struct dst_entry *dst, *next, head;
  71         struct dst_entry *last = &head;
  72 #if RT_CACHE_DEBUG >= 2
  73         ktime_t time_start = ktime_get();
  74         struct timespec elapsed;
  75 #endif
  76
  77         mutex_lock(&dst_gc_mutex);
  78         next = dst_busy_list;
  79
  80 loop:
  81         while ((dst = next) != NULL) {
  82                 next = dst->next;
  83                 prefetch(&next->next);
  84                 cond_resched();
  85                 if (likely(atomic_read(&dst->__refcnt))) {
  86                         last->next = dst;
  87                         last = dst;
  88                         delayed++;
  89                         continue;
  90                 }
  91                 work_performed++;
  92
  93                 dst = dst_destroy(dst);
  94                 if (dst) {
  95                         /* NOHASH and still referenced. Unless it is already
  96                          * on gc list, invalidate it and add to gc list.
  97                          *
  98                          * Note: this is temporary. Actually, NOHASH dst's
  99                          * must be obsoleted when parent is obsoleted.
 100                          * But we do not have state "obsoleted, but
 101                          * referenced by parent", so it is right.
 102                          */
 103                         if (dst->obsolete > 1)
 104                                 continue;
 105
 106                         ___dst_free(dst);
 107                         dst->next = next;
 108                         next = dst;
 109                 }
 110         }
 111
 112         spin_lock_bh(&dst_garbage.lock);
 113         next = dst_garbage.list;
 114         if (next) {
 115                 dst_garbage.list = NULL;
 116                 spin_unlock_bh(&dst_garbage.lock);
 117                 goto loop;
 118         }
 119         last->next = NULL;
 120         dst_busy_list = head.next;
 121         if (!dst_busy_list)
 122                 dst_garbage.timer_inc = DST_GC_MAX;
 123         else {
 124                 /*
 125                  * if we freed less than 1/10 of delayed entries,
 126                  * we can sleep longer.
 127                  */
 128                 if (work_performed <= delayed/10) {
 129                         dst_garbage.timer_expires += dst_garbage.timer_inc;
 130                         if (dst_garbage.timer_expires > DST_GC_MAX)
 131                                 dst_garbage.timer_expires = DST_GC_MAX;
 132                         dst_garbage.timer_inc += DST_GC_INC;
 133                 } else {
 134                         dst_garbage.timer_inc = DST_GC_INC;
 135                         dst_garbage.timer_expires = DST_GC_MIN;
 136                 }
 137                 expires = dst_garbage.timer_expires;
 138                 /*
 139                  * if the next desired timer is more than 4 seconds in the
 140                  * future then round the timer to whole seconds
 141                  */
 142                 if (expires > 4*HZ)
 143                         expires = round_jiffies_relative(expires);
 144                 schedule_delayed_work(&dst_gc_work, expires);
 145         }
 146
 147         spin_unlock_bh(&dst_garbage.lock);
 148         mutex_unlock(&dst_gc_mutex);
 149 #if RT_CACHE_DEBUG >= 2
 150         elapsed = ktime_to_timespec(ktime_sub(ktime_get(), time_start));
 151         printk(KERN_DEBUG "dst_total: %d delayed: %d work_perf: %d"
 152                 " expires: %lu elapsed: %lu us\n",
 153                 atomic_read(&dst_total), delayed, work_performed,
 154                 expires,
 155                 elapsed.tv_sec * USEC_PER_SEC +
 156                   elapsed.tv_nsec / NSEC_PER_USEC);
 157 #endif
 158 }
 159
 160 int dst_discard(struct sk_buff *skb)
 161 {
 162         kfree_skb(skb);
 163         return 0;
 164 }
 165 EXPORT_SYMBOL(dst_discard);
 166
 167 const u32 dst_default_metrics[RTAX_MAX];
 168
 169 void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
 170                 int initial_ref, int initial_obsolete, int flags)
 171 {
 172         struct dst_entry *dst;
 173
 174         if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
 175                 if (ops->gc(ops))
 176                         return NULL;
 177         }
 178         dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
 179         if (!dst)
 180                 return NULL;
 181         dst->child = NULL;
 182         dst->dev = dev;
 183         if (dev)
 184                 dev_hold(dev);
 185         dst->ops = ops;
 186         dst_init_metrics(dst, dst_default_metrics, true);
 187         dst->expires = 0UL;
 188         dst->path = dst;
 189         dst->neighbour = NULL;
 190         dst->hh = NULL;
 191 #ifdef CONFIG_XFRM
 192         dst->xfrm = NULL;
 193 #endif
 194         dst->input = dst_discard;
 195         dst->output = dst_discard;
 196         dst->error = 0;
 197         dst->obsolete = initial_obsolete;
 198         dst->header_len = 0;
 199         dst->trailer_len = 0;
 200 #ifdef CONFIG_IP_ROUTE_CLASSID
 201         dst->tclassid = 0;
 202 #endif
 203         atomic_set(&dst->__refcnt, initial_ref);
 204         dst->__use = 0;
 205         dst->lastuse = jiffies;
 206         dst->flags = flags;
 207         dst->next = NULL;
 208 #if RT_CACHE_DEBUG >= 2
 209         atomic_inc(&dst_total);
 210 #endif
 211         dst_entries_add(ops, 1);
 212         return dst;
 213 }
 214 EXPORT_SYMBOL(dst_alloc);
 215
 216 static void ___dst_free(struct dst_entry *dst)
 217 {
 218         /* The first case (dev==NULL) is required, when
 219            protocol module is unloaded.
 220          */
 221         if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
 222                 dst->input = dst->output = dst_discard;
 223         dst->obsolete = 2;
 224 }
 225
 226 void __dst_free(struct dst_entry *dst)
 227 {
 228         spin_lock_bh(&dst_garbage.lock);
 229         ___dst_free(dst);
 230         dst->next = dst_garbage.list;
 231         dst_garbage.list = dst;
 232         if (dst_garbage.timer_inc > DST_GC_INC) {
 233                 dst_garbage.timer_inc = DST_GC_INC;
 234                 dst_garbage.timer_expires = DST_GC_MIN;
 235                 cancel_delayed_work(&dst_gc_work);
 236                 schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
 237         }
 238         spin_unlock_bh(&dst_garbage.lock);
 239 }
 240 EXPORT_SYMBOL(__dst_free);
 241
 242 struct dst_entry *dst_destroy(struct dst_entry * dst)
 243 {
 244         struct dst_entry *child;
 245         struct neighbour *neigh;
 246         struct hh_cache *hh;
 247
 248         smp_rmb();
 249
 250 again:
 251         neigh = dst->neighbour;
 252         hh = dst->hh;
 253         child = dst->child;
 254
 255         dst->hh = NULL;
 256         if (hh)
 257                 hh_cache_put(hh);
 258
 259         if (neigh) {
 260                 dst->neighbour = NULL;
 261                 neigh_release(neigh);
 262         }
 263
 264         dst_entries_add(dst->ops, -1);
 265
 266         if (dst->ops->destroy)
 267                 dst->ops->destroy(dst);
 268         if (dst->dev)
 269                 dev_put(dst->dev);
 270 #if RT_CACHE_DEBUG >= 2
 271         atomic_dec(&dst_total);
 272 #endif
 273         kmem_cache_free(dst->ops->kmem_cachep, dst);
 274
 275         dst = child;
 276         if (dst) {
 277                 int nohash = dst->flags & DST_NOHASH;
 278
 279                 if (atomic_dec_and_test(&dst->__refcnt)) {
 280                         /* We were real parent of this dst, so kill child. */
 281                         if (nohash)
 282                                 goto again;
 283                 } else {
 284                         /* Child is still referenced, return it for freeing. */
 285                         if (nohash)
 286                                 return dst;
 287                         /* Child is still in his hash table */
 288                 }
 289         }
 290         return NULL;
 291 }
 292 EXPORT_SYMBOL(dst_destroy);
 293
 294 void dst_release(struct dst_entry *dst)
 295 {
 296         if (dst) {
 297                 int newrefcnt;
 298
 299                 newrefcnt = atomic_dec_return(&dst->__refcnt);
 300                 WARN_ON(newrefcnt < 0);
 301                 if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
 302                         dst = dst_destroy(dst);
 303                         if (dst)
 304                                 __dst_free(dst);
 305                 }
 306         }
 307 }
 308 EXPORT_SYMBOL(dst_release);
 309
 310 u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
 311 {
 312         u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC);
 313
 314         if (p) {
 315                 u32 *old_p = __DST_METRICS_PTR(old);
 316                 unsigned long prev, new;
 317
 318                 memcpy(p, old_p, sizeof(u32) * RTAX_MAX);
 319
 320                 new = (unsigned long) p;
 321                 prev = cmpxchg(&dst->_metrics, old, new);
 322
 323                 if (prev != old) {
 324                         kfree(p);
 325                         p = __DST_METRICS_PTR(prev);
 326                         if (prev & DST_METRICS_READ_ONLY)
 327                                 p = NULL;
 328                 }
 329         }
 330         return p;
 331 }
 332 EXPORT_SYMBOL(dst_cow_metrics_generic);
 333
 334 /* Caller asserts that dst_metrics_read_only(dst) is false.  */
 335 void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
 336 {
 337         unsigned long prev, new;
 338
 339         new = (unsigned long) dst_default_metrics;
 340         prev = cmpxchg(&dst->_metrics, old, new);
 341         if (prev == old)
 342                 kfree(__DST_METRICS_PTR(old));
 343 }
 344 EXPORT_SYMBOL(__dst_destroy_metrics_generic);
 345
 346 /**
 347  * skb_dst_set_noref - sets skb dst, without a reference
 348  * @skb: buffer
 349  * @dst: dst entry
 350  *
 351  * Sets skb dst, assuming a reference was not taken on dst
 352  * skb_dst_drop() should not dst_release() this dst
 353  */
 354 void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
 355 {
 356         WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
 357         /* If dst not in cache, we must take a reference, because
 358          * dst_release() will destroy dst as soon as its refcount becomes zero
 359          */
 360         if (unlikely(dst->flags & DST_NOCACHE)) {
 361                 dst_hold(dst);
 362                 skb_dst_set(skb, dst);
 363         } else {
 364                 skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
 365         }
 366 }
 367 EXPORT_SYMBOL(skb_dst_set_noref);
 368
 369 /* Dirty hack. We did it in 2.2 (in __dst_free),
 370  * we have _very_ good reasons not to repeat
 371  * this mistake in 2.3, but we have no choice
 372  * now. _It_ _is_ _explicit_ _deliberate_
 373  * _race_ _condition_.
 374  *
 375  * Commented and originally written by Alexey.
 376  */
 377 static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 378                        int unregister)
 379 {
 380         if (dst->ops->ifdown)
 381                 dst->ops->ifdown(dst, dev, unregister);
 382
 383         if (dev != dst->dev)
 384                 return;
 385
 386         if (!unregister) {
 387                 dst->input = dst->output = dst_discard;
 388         } else {
 389                 dst->dev = dev_net(dst->dev)->loopback_dev;
 390                 dev_hold(dst->dev);
 391                 dev_put(dev);
 392                 if (dst->neighbour && dst->neighbour->dev == dev) {
 393                         dst->neighbour->dev = dst->dev;
 394                         dev_hold(dst->dev);
 395                         dev_put(dev);
 396                 }
 397         }
 398 }
 399
 400 static int dst_dev_event(struct notifier_block *this, unsigned long event,
 401                          void *ptr)
 402 {
 403         struct net_device *dev = ptr;
 404         struct dst_entry *dst, *last = NULL;
 405
 406         switch (event) {
 407         case NETDEV_UNREGISTER:
 408         case NETDEV_DOWN:
 409                 mutex_lock(&dst_gc_mutex);
 410                 for (dst = dst_busy_list; dst; dst = dst->next) {
 411                         last = dst;
 412                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 413                 }
 414
 415                 spin_lock_bh(&dst_garbage.lock);
 416                 dst = dst_garbage.list;
 417                 dst_garbage.list = NULL;
 418                 spin_unlock_bh(&dst_garbage.lock);
 419
 420                 if (last)
 421                         last->next = dst;
 422                 else
 423                         dst_busy_list = dst;
 424                 for (; dst; dst = dst->next)
 425                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 426                 mutex_unlock(&dst_gc_mutex);
 427                 break;
 428         }
 429         return NOTIFY_DONE;
 430 }
 431
 432 static struct notifier_block dst_dev_notifier = {
 433         .notifier_call  = dst_dev_event,
 434         .priority = -10, /* must be called after other network notifiers */
 435 };
 436
 437 void __init dst_init(void)
 438 {
 439         register_netdevice_notifier(&dst_dev_notifier);
 440 }