[NET]: dev_put/dev_hold cleanup
[linux-2.6.git] / net / ipv4 / inet_timewait_sock.c
index d38d160..417f126 100644 (file)
@@ -12,6 +12,7 @@
 
 #include <net/inet_hashtables.h>
 #include <net/inet_timewait_sock.h>
+#include <net/ip.h>
 
 /* Must be called with locally disabled BHs. */
 void __inet_twsk_kill(struct inet_timewait_sock *tw, struct inet_hashinfo *hashinfo)
@@ -19,7 +20,7 @@ void __inet_twsk_kill(struct inet_timewait_sock *tw, struct inet_hashinfo *hashi
        struct inet_bind_hashbucket *bhead;
        struct inet_bind_bucket *tb;
        /* Unlink from established hashes. */
-       struct inet_ehash_bucket *ehead = &hashinfo->ehash[tw->tw_hashent];
+       struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, tw->tw_hash);
 
        write_lock(&ehead->lock);
        if (hlist_unhashed(&tw->tw_node)) {
@@ -47,6 +48,8 @@ void __inet_twsk_kill(struct inet_timewait_sock *tw, struct inet_hashinfo *hashi
        inet_twsk_put(tw);
 }
 
+EXPORT_SYMBOL_GPL(__inet_twsk_kill);
+
 /*
  * Enter the time wait state. This is called with locally disabled BH.
  * Essentially we whip up a timewait bucket, copy the relevant info into it
@@ -56,7 +59,8 @@ void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
                           struct inet_hashinfo *hashinfo)
 {
        const struct inet_sock *inet = inet_sk(sk);
-       struct inet_ehash_bucket *ehead = &hashinfo->ehash[sk->sk_hashent];
+       const struct inet_connection_sock *icsk = inet_csk(sk);
+       struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, sk->sk_hash);
        struct inet_bind_hashbucket *bhead;
        /* Step 1: Put TW into bind hash. Original socket stays there too.
           Note, that any socket with inet->num != 0 MUST be bound in
@@ -64,8 +68,8 @@ void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
         */
        bhead = &hashinfo->bhash[inet_bhashfn(inet->num, hashinfo->bhash_size)];
        spin_lock(&bhead->lock);
-       tw->tw_tb = inet->bind_hash;
-       BUG_TRAP(inet->bind_hash);
+       tw->tw_tb = icsk->icsk_bind_hash;
+       BUG_TRAP(icsk->icsk_bind_hash);
        inet_twsk_add_bind_node(tw, &tw->tw_tb->owners);
        spin_unlock(&bhead->lock);
 
@@ -81,3 +85,302 @@ void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
 
        write_unlock(&ehead->lock);
 }
+
+EXPORT_SYMBOL_GPL(__inet_twsk_hashdance);
+
+struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk, const int state)
+{
+       struct inet_timewait_sock *tw =
+               kmem_cache_alloc(sk->sk_prot_creator->twsk_prot->twsk_slab,
+                                SLAB_ATOMIC);
+       if (tw != NULL) {
+               const struct inet_sock *inet = inet_sk(sk);
+
+               /* Give us an identity. */
+               tw->tw_daddr        = inet->daddr;
+               tw->tw_rcv_saddr    = inet->rcv_saddr;
+               tw->tw_bound_dev_if = sk->sk_bound_dev_if;
+               tw->tw_num          = inet->num;
+               tw->tw_state        = TCP_TIME_WAIT;
+               tw->tw_substate     = state;
+               tw->tw_sport        = inet->sport;
+               tw->tw_dport        = inet->dport;
+               tw->tw_family       = sk->sk_family;
+               tw->tw_reuse        = sk->sk_reuse;
+               tw->tw_hash         = sk->sk_hash;
+               tw->tw_ipv6only     = 0;
+               tw->tw_prot         = sk->sk_prot_creator;
+               atomic_set(&tw->tw_refcnt, 1);
+               inet_twsk_dead_node_init(tw);
+               __module_get(tw->tw_prot->owner);
+       }
+
+       return tw;
+}
+
+EXPORT_SYMBOL_GPL(inet_twsk_alloc);
+
+/* Returns non-zero if quota exceeded.  */
+static int inet_twdr_do_twkill_work(struct inet_timewait_death_row *twdr,
+                                   const int slot)
+{
+       struct inet_timewait_sock *tw;
+       struct hlist_node *node;
+       unsigned int killed;
+       int ret;
+
+       /* NOTE: compare this to previous version where lock
+        * was released after detaching chain. It was racy,
+        * because tw buckets are scheduled in not serialized context
+        * in 2.3 (with netfilter), and with softnet it is common, because
+        * soft irqs are not sequenced.
+        */
+       killed = 0;
+       ret = 0;
+rescan:
+       inet_twsk_for_each_inmate(tw, node, &twdr->cells[slot]) {
+               __inet_twsk_del_dead_node(tw);
+               spin_unlock(&twdr->death_lock);
+               __inet_twsk_kill(tw, twdr->hashinfo);
+               inet_twsk_put(tw);
+               killed++;
+               spin_lock(&twdr->death_lock);
+               if (killed > INET_TWDR_TWKILL_QUOTA) {
+                       ret = 1;
+                       break;
+               }
+
+               /* While we dropped twdr->death_lock, another cpu may have
+                * killed off the next TW bucket in the list, therefore
+                * do a fresh re-read of the hlist head node with the
+                * lock reacquired.  We still use the hlist traversal
+                * macro in order to get the prefetches.
+                */
+               goto rescan;
+       }
+
+       twdr->tw_count -= killed;
+       NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITED, killed);
+
+       return ret;
+}
+
+void inet_twdr_hangman(unsigned long data)
+{
+       struct inet_timewait_death_row *twdr;
+       int unsigned need_timer;
+
+       twdr = (struct inet_timewait_death_row *)data;
+       spin_lock(&twdr->death_lock);
+
+       if (twdr->tw_count == 0)
+               goto out;
+
+       need_timer = 0;
+       if (inet_twdr_do_twkill_work(twdr, twdr->slot)) {
+               twdr->thread_slots |= (1 << twdr->slot);
+               mb();
+               schedule_work(&twdr->twkill_work);
+               need_timer = 1;
+       } else {
+               /* We purged the entire slot, anything left?  */
+               if (twdr->tw_count)
+                       need_timer = 1;
+       }
+       twdr->slot = ((twdr->slot + 1) & (INET_TWDR_TWKILL_SLOTS - 1));
+       if (need_timer)
+               mod_timer(&twdr->tw_timer, jiffies + twdr->period);
+out:
+       spin_unlock(&twdr->death_lock);
+}
+
+EXPORT_SYMBOL_GPL(inet_twdr_hangman);
+
+extern void twkill_slots_invalid(void);
+
+void inet_twdr_twkill_work(void *data)
+{
+       struct inet_timewait_death_row *twdr = data;
+       int i;
+
+       if ((INET_TWDR_TWKILL_SLOTS - 1) > (sizeof(twdr->thread_slots) * 8))
+               twkill_slots_invalid();
+
+       while (twdr->thread_slots) {
+               spin_lock_bh(&twdr->death_lock);
+               for (i = 0; i < INET_TWDR_TWKILL_SLOTS; i++) {
+                       if (!(twdr->thread_slots & (1 << i)))
+                               continue;
+
+                       while (inet_twdr_do_twkill_work(twdr, i) != 0) {
+                               if (need_resched()) {
+                                       spin_unlock_bh(&twdr->death_lock);
+                                       schedule();
+                                       spin_lock_bh(&twdr->death_lock);
+                               }
+                       }
+
+                       twdr->thread_slots &= ~(1 << i);
+               }
+               spin_unlock_bh(&twdr->death_lock);
+       }
+}
+
+EXPORT_SYMBOL_GPL(inet_twdr_twkill_work);
+
+/* These are always called from BH context.  See callers in
+ * tcp_input.c to verify this.
+ */
+
+/* This is for handling early-kills of TIME_WAIT sockets. */
+void inet_twsk_deschedule(struct inet_timewait_sock *tw,
+                         struct inet_timewait_death_row *twdr)
+{
+       spin_lock(&twdr->death_lock);
+       if (inet_twsk_del_dead_node(tw)) {
+               inet_twsk_put(tw);
+               if (--twdr->tw_count == 0)
+                       del_timer(&twdr->tw_timer);
+       }
+       spin_unlock(&twdr->death_lock);
+       __inet_twsk_kill(tw, twdr->hashinfo);
+}
+
+EXPORT_SYMBOL(inet_twsk_deschedule);
+
+void inet_twsk_schedule(struct inet_timewait_sock *tw,
+                      struct inet_timewait_death_row *twdr,
+                      const int timeo, const int timewait_len)
+{
+       struct hlist_head *list;
+       int slot;
+
+       /* timeout := RTO * 3.5
+        *
+        * 3.5 = 1+2+0.5 to wait for two retransmits.
+        *
+        * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
+        * our ACK acking that FIN can be lost. If N subsequent retransmitted
+        * FINs (or previous seqments) are lost (probability of such event
+        * is p^(N+1), where p is probability to lose single packet and
+        * time to detect the loss is about RTO*(2^N - 1) with exponential
+        * backoff). Normal timewait length is calculated so, that we
+        * waited at least for one retransmitted FIN (maximal RTO is 120sec).
+        * [ BTW Linux. following BSD, violates this requirement waiting
+        *   only for 60sec, we should wait at least for 240 secs.
+        *   Well, 240 consumes too much of resources 8)
+        * ]
+        * This interval is not reduced to catch old duplicate and
+        * responces to our wandering segments living for two MSLs.
+        * However, if we use PAWS to detect
+        * old duplicates, we can reduce the interval to bounds required
+        * by RTO, rather than MSL. So, if peer understands PAWS, we
+        * kill tw bucket after 3.5*RTO (it is important that this number
+        * is greater than TS tick!) and detect old duplicates with help
+        * of PAWS.
+        */
+       slot = (timeo + (1 << INET_TWDR_RECYCLE_TICK) - 1) >> INET_TWDR_RECYCLE_TICK;
+
+       spin_lock(&twdr->death_lock);
+
+       /* Unlink it, if it was scheduled */
+       if (inet_twsk_del_dead_node(tw))
+               twdr->tw_count--;
+       else
+               atomic_inc(&tw->tw_refcnt);
+
+       if (slot >= INET_TWDR_RECYCLE_SLOTS) {
+               /* Schedule to slow timer */
+               if (timeo >= timewait_len) {
+                       slot = INET_TWDR_TWKILL_SLOTS - 1;
+               } else {
+                       slot = (timeo + twdr->period - 1) / twdr->period;
+                       if (slot >= INET_TWDR_TWKILL_SLOTS)
+                               slot = INET_TWDR_TWKILL_SLOTS - 1;
+               }
+               tw->tw_ttd = jiffies + timeo;
+               slot = (twdr->slot + slot) & (INET_TWDR_TWKILL_SLOTS - 1);
+               list = &twdr->cells[slot];
+       } else {
+               tw->tw_ttd = jiffies + (slot << INET_TWDR_RECYCLE_TICK);
+
+               if (twdr->twcal_hand < 0) {
+                       twdr->twcal_hand = 0;
+                       twdr->twcal_jiffie = jiffies;
+                       twdr->twcal_timer.expires = twdr->twcal_jiffie +
+                                             (slot << INET_TWDR_RECYCLE_TICK);
+                       add_timer(&twdr->twcal_timer);
+               } else {
+                       if (time_after(twdr->twcal_timer.expires,
+                                      jiffies + (slot << INET_TWDR_RECYCLE_TICK)))
+                               mod_timer(&twdr->twcal_timer,
+                                         jiffies + (slot << INET_TWDR_RECYCLE_TICK));
+                       slot = (twdr->twcal_hand + slot) & (INET_TWDR_RECYCLE_SLOTS - 1);
+               }
+               list = &twdr->twcal_row[slot];
+       }
+
+       hlist_add_head(&tw->tw_death_node, list);
+
+       if (twdr->tw_count++ == 0)
+               mod_timer(&twdr->tw_timer, jiffies + twdr->period);
+       spin_unlock(&twdr->death_lock);
+}
+
+EXPORT_SYMBOL_GPL(inet_twsk_schedule);
+
+void inet_twdr_twcal_tick(unsigned long data)
+{
+       struct inet_timewait_death_row *twdr;
+       int n, slot;
+       unsigned long j;
+       unsigned long now = jiffies;
+       int killed = 0;
+       int adv = 0;
+
+       twdr = (struct inet_timewait_death_row *)data;
+
+       spin_lock(&twdr->death_lock);
+       if (twdr->twcal_hand < 0)
+               goto out;
+
+       slot = twdr->twcal_hand;
+       j = twdr->twcal_jiffie;
+
+       for (n = 0; n < INET_TWDR_RECYCLE_SLOTS; n++) {
+               if (time_before_eq(j, now)) {
+                       struct hlist_node *node, *safe;
+                       struct inet_timewait_sock *tw;
+
+                       inet_twsk_for_each_inmate_safe(tw, node, safe,
+                                                      &twdr->twcal_row[slot]) {
+                               __inet_twsk_del_dead_node(tw);
+                               __inet_twsk_kill(tw, twdr->hashinfo);
+                               inet_twsk_put(tw);
+                               killed++;
+                       }
+               } else {
+                       if (!adv) {
+                               adv = 1;
+                               twdr->twcal_jiffie = j;
+                               twdr->twcal_hand = slot;
+                       }
+
+                       if (!hlist_empty(&twdr->twcal_row[slot])) {
+                               mod_timer(&twdr->twcal_timer, j);
+                               goto out;
+                       }
+               }
+               j += 1 << INET_TWDR_RECYCLE_TICK;
+               slot = (slot + 1) & (INET_TWDR_RECYCLE_SLOTS - 1);
+       }
+       twdr->twcal_hand = -1;
+
+out:
+       if ((twdr->tw_count -= killed) == 0)
+               del_timer(&twdr->tw_timer);
+       NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITKILLED, killed);
+       spin_unlock(&twdr->death_lock);
+}
+
+EXPORT_SYMBOL_GPL(inet_twdr_twcal_tick);