linux/net/core/dst.c
<<
>>
Prefs
   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#include <linux/prefetch.h>
  23
  24#include <net/dst.h>
  25
  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.
  36 */
  37
  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 */
  43static struct {
  44        spinlock_t              lock;
  45        struct dst_entry        *list;
  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};
  52static void dst_gc_task(struct work_struct *work);
  53static void ___dst_free(struct dst_entry *dst);
  54
  55static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
  56
  57static DEFINE_MUTEX(dst_gc_mutex);
  58/*
  59 * long lived entries are maintained in this list, guarded by dst_gc_mutex
  60 */
  61static struct dst_entry         *dst_busy_list;
  62
  63static void dst_gc_task(struct work_struct *work)
  64{
  65        int    delayed = 0;
  66        int    work_performed = 0;
  67        unsigned long expires = ~0L;
  68        struct dst_entry *dst, *next, head;
  69        struct dst_entry *last = &head;
  70
  71        mutex_lock(&dst_gc_mutex);
  72        next = dst_busy_list;
  73
  74loop:
  75        while ((dst = next) != NULL) {
  76                next = dst->next;
  77                prefetch(&next->next);
  78                cond_resched();
  79                if (likely(atomic_read(&dst->__refcnt))) {
  80                        last->next = dst;
  81                        last = dst;
  82                        delayed++;
  83                        continue;
  84                }
  85                work_performed++;
  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                         */
  97                        if (dst->obsolete > 0)
  98                                continue;
  99
 100                        ___dst_free(dst);
 101                        dst->next = next;
 102                        next = dst;
 103                }
 104        }
 105
 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                /*
 133                 * if the next desired timer is more than 4 seconds in the
 134                 * future then round the timer to whole seconds
 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);
 143}
 144
 145int dst_discard(struct sk_buff *skb)
 146{
 147        kfree_skb(skb);
 148        return 0;
 149}
 150EXPORT_SYMBOL(dst_discard);
 151
 152const 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
 161
 162void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
 163                int initial_ref, int initial_obsolete, unsigned short flags)
 164{
 165        struct dst_entry *dst;
 166
 167        if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
 168                if (ops->gc(ops))
 169                        return NULL;
 170        }
 171        dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
 172        if (!dst)
 173                return NULL;
 174        dst->child = NULL;
 175        dst->dev = dev;
 176        if (dev)
 177                dev_hold(dev);
 178        dst->ops = ops;
 179        dst_init_metrics(dst, dst_default_metrics, true);
 180        dst->expires = 0UL;
 181        dst->path = dst;
 182#ifdef CONFIG_XFRM
 183        dst->xfrm = NULL;
 184#endif
 185        dst->input = dst_discard;
 186        dst->output = dst_discard;
 187        dst->error = 0;
 188        dst->obsolete = initial_obsolete;
 189        dst->header_len = 0;
 190        dst->trailer_len = 0;
 191#ifdef CONFIG_IP_ROUTE_CLASSID
 192        dst->tclassid = 0;
 193#endif
 194        atomic_set(&dst->__refcnt, initial_ref);
 195        dst->__use = 0;
 196        dst->lastuse = jiffies;
 197        dst->flags = flags;
 198        dst->pending_confirm = 0;
 199        dst->next = NULL;
 200        if (!(flags & DST_NOCOUNT))
 201                dst_entries_add(ops, 1);
 202        return dst;
 203}
 204EXPORT_SYMBOL(dst_alloc);
 205
 206static void ___dst_free(struct dst_entry *dst)
 207{
 208        /* The first case (dev==NULL) is required, when
 209           protocol module is unloaded.
 210         */
 211        if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
 212                dst->input = dst->output = dst_discard;
 213        dst->obsolete = DST_OBSOLETE_DEAD;
 214}
 215
 216void __dst_free(struct dst_entry *dst)
 217{
 218        spin_lock_bh(&dst_garbage.lock);
 219        ___dst_free(dst);
 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;
 225                cancel_delayed_work(&dst_gc_work);
 226                schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
 227        }
 228        spin_unlock_bh(&dst_garbage.lock);
 229}
 230EXPORT_SYMBOL(__dst_free);
 231
 232struct dst_entry *dst_destroy(struct dst_entry * dst)
 233{
 234        struct dst_entry *child;
 235
 236        smp_rmb();
 237
 238again:
 239        child = dst->child;
 240
 241        if (!(dst->flags & DST_NOCOUNT))
 242                dst_entries_add(dst->ops, -1);
 243
 244        if (dst->ops->destroy)
 245                dst->ops->destroy(dst);
 246        if (dst->dev)
 247                dev_put(dst->dev);
 248        kmem_cache_free(dst->ops->kmem_cachep, dst);
 249
 250        dst = child;
 251        if (dst) {
 252                int nohash = dst->flags & DST_NOHASH;
 253
 254                if (atomic_dec_and_test(&dst->__refcnt)) {
 255                        /* We were real parent of this dst, so kill child. */
 256                        if (nohash)
 257                                goto again;
 258                } else {
 259                        /* Child is still referenced, return it for freeing. */
 260                        if (nohash)
 261                                return dst;
 262                        /* Child is still in his hash table */
 263                }
 264        }
 265        return NULL;
 266}
 267EXPORT_SYMBOL(dst_destroy);
 268
 269void dst_release(struct dst_entry *dst)
 270{
 271        if (dst) {
 272                int newrefcnt;
 273
 274                newrefcnt = atomic_dec_return(&dst->__refcnt);
 275                WARN_ON(newrefcnt < 0);
 276                if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
 277                        dst = dst_destroy(dst);
 278                        if (dst)
 279                                __dst_free(dst);
 280                }
 281        }
 282}
 283EXPORT_SYMBOL(dst_release);
 284
 285u32 *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}
 307EXPORT_SYMBOL(dst_cow_metrics_generic);
 308
 309/* Caller asserts that dst_metrics_read_only(dst) is false.  */
 310void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
 311{
 312        unsigned long prev, new;
 313
 314        new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY;
 315        prev = cmpxchg(&dst->_metrics, old, new);
 316        if (prev == old)
 317                kfree(__DST_METRICS_PTR(old));
 318}
 319EXPORT_SYMBOL(__dst_destroy_metrics_generic);
 320
 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 */
 329void 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}
 342EXPORT_SYMBOL(skb_dst_set_noref);
 343
 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 */
 352static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 353                       int unregister)
 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) {
 362                dst->input = dst->output = dst_discard;
 363        } else {
 364                dst->dev = dev_net(dst->dev)->loopback_dev;
 365                dev_hold(dst->dev);
 366                dev_put(dev);
 367        }
 368}
 369
 370static int dst_dev_event(struct notifier_block *this, unsigned long event,
 371                         void *ptr)
 372{
 373        struct net_device *dev = ptr;
 374        struct dst_entry *dst, *last = NULL;
 375
 376        switch (event) {
 377        case NETDEV_UNREGISTER:
 378        case NETDEV_DOWN:
 379                mutex_lock(&dst_gc_mutex);
 380                for (dst = dst_busy_list; dst; dst = dst->next) {
 381                        last = dst;
 382                        dst_ifdown(dst, dev, event != NETDEV_DOWN);
 383                }
 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;
 394                for (; dst; dst = dst->next)
 395                        dst_ifdown(dst, dev, event != NETDEV_DOWN);
 396                mutex_unlock(&dst_gc_mutex);
 397                break;
 398        }
 399        return NOTIFY_DONE;
 400}
 401
 402static struct notifier_block dst_dev_notifier = {
 403        .notifier_call  = dst_dev_event,
 404        .priority = -10, /* must be called after other network notifiers */
 405};
 406
 407void __init dst_init(void)
 408{
 409        register_netdevice_notifier(&dst_dev_notifier);
 410}
 411
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.