aio: don't include aio.h in sched.h
[linux-3.10.git] / net / rds / bind.c
index c17cc39..b5ad65a 100644 (file)
 #include <net/sock.h>
 #include <linux/in.h>
 #include <linux/if_arp.h>
+#include <linux/jhash.h>
+#include <linux/ratelimit.h>
 #include "rds.h"
 
-/*
- * XXX this probably still needs more work.. no INADDR_ANY, and rbtrees aren't
- * particularly zippy.
- *
- * This is now called for every incoming frame so we arguably care much more
- * about it than we used to.
- */
+#define BIND_HASH_SIZE 1024
+static struct hlist_head bind_hash_table[BIND_HASH_SIZE];
 static DEFINE_SPINLOCK(rds_bind_lock);
-static struct rb_root rds_bind_tree = RB_ROOT;
 
-static struct rds_sock *rds_bind_tree_walk(__be32 addr, __be16 port,
-                                          struct rds_sock *insert)
+static struct hlist_head *hash_to_bucket(__be32 addr, __be16 port)
+{
+       return bind_hash_table + (jhash_2words((u32)addr, (u32)port, 0) &
+                                 (BIND_HASH_SIZE - 1));
+}
+
+static struct rds_sock *rds_bind_lookup(__be32 addr, __be16 port,
+                                       struct rds_sock *insert)
 {
-       struct rb_node **p = &rds_bind_tree.rb_node;
-       struct rb_node *parent = NULL;
        struct rds_sock *rs;
+       struct hlist_head *head = hash_to_bucket(addr, port);
        u64 cmp;
        u64 needle = ((u64)be32_to_cpu(addr) << 32) | be16_to_cpu(port);
 
-       while (*p) {
-               parent = *p;
-               rs = rb_entry(parent, struct rds_sock, rs_bound_node);
-
+       rcu_read_lock();
+       hlist_for_each_entry_rcu(rs, head, rs_bound_node) {
                cmp = ((u64)be32_to_cpu(rs->rs_bound_addr) << 32) |
                      be16_to_cpu(rs->rs_bound_port);
 
-               if (needle < cmp)
-                       p = &(*p)->rb_left;
-               else if (needle > cmp)
-                       p = &(*p)->rb_right;
-               else
+               if (cmp == needle) {
+                       rcu_read_unlock();
                        return rs;
+               }
        }
+       rcu_read_unlock();
 
        if (insert) {
-               rb_link_node(&insert->rs_bound_node, parent, p);
-               rb_insert_color(&insert->rs_bound_node, &rds_bind_tree);
+               /*
+                * make sure our addr and port are set before
+                * we are added to the list, other people
+                * in rcu will find us as soon as the
+                * hlist_add_head_rcu is done
+                */
+               insert->rs_bound_addr = addr;
+               insert->rs_bound_port = port;
+               rds_sock_addref(insert);
+
+               hlist_add_head_rcu(&insert->rs_bound_node, head);
        }
        return NULL;
 }
@@ -86,15 +93,13 @@ static struct rds_sock *rds_bind_tree_walk(__be32 addr, __be16 port,
 struct rds_sock *rds_find_bound(__be32 addr, __be16 port)
 {
        struct rds_sock *rs;
-       unsigned long flags;
 
-       spin_lock_irqsave(&rds_bind_lock, flags);
-       rs = rds_bind_tree_walk(addr, port, NULL);
+       rs = rds_bind_lookup(addr, port, NULL);
+
        if (rs && !sock_flag(rds_rs_to_sk(rs), SOCK_DEAD))
                rds_sock_addref(rs);
        else
                rs = NULL;
-       spin_unlock_irqrestore(&rds_bind_lock, flags);
 
        rdsdebug("returning rs %p for %pI4:%u\n", rs, &addr,
                ntohs(port));
@@ -121,22 +126,15 @@ static int rds_add_bound(struct rds_sock *rs, __be32 addr, __be16 *port)
        do {
                if (rover == 0)
                        rover++;
-               if (rds_bind_tree_walk(addr, cpu_to_be16(rover), rs) == NULL) {
-                       *port = cpu_to_be16(rover);
+               if (!rds_bind_lookup(addr, cpu_to_be16(rover), rs)) {
+                       *port = rs->rs_bound_port;
                        ret = 0;
+                       rdsdebug("rs %p binding to %pI4:%d\n",
+                         rs, &addr, (int)ntohs(*port));
                        break;
                }
        } while (rover++ != last);
 
-       if (ret == 0)  {
-               rs->rs_bound_addr = addr;
-               rs->rs_bound_port = *port;
-               rds_sock_addref(rs);
-
-               rdsdebug("rs %p binding to %pI4:%d\n",
-                 rs, &addr, (int)ntohs(*port));
-       }
-
        spin_unlock_irqrestore(&rds_bind_lock, flags);
 
        return ret;
@@ -153,7 +151,7 @@ void rds_remove_bound(struct rds_sock *rs)
                  rs, &rs->rs_bound_addr,
                  ntohs(rs->rs_bound_port));
 
-               rb_erase(&rs->rs_bound_node, &rds_bind_tree);
+               hlist_del_init_rcu(&rs->rs_bound_node);
                rds_sock_put(rs);
                rs->rs_bound_addr = 0;
        }
@@ -184,9 +182,11 @@ int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
                goto out;
 
        trans = rds_trans_get_preferred(sin->sin_addr.s_addr);
-       if (trans == NULL) {
+       if (!trans) {
                ret = -EADDRNOTAVAIL;
                rds_remove_bound(rs);
+               printk_ratelimited(KERN_INFO "RDS: rds_bind() could not find a transport, "
+                               "load rds_tcp or rds_rdma?\n");
                goto out;
        }
 
@@ -195,5 +195,9 @@ int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
 
 out:
        release_sock(sk);
+
+       /* we might have called rds_remove_bound on error */
+       if (ret)
+               synchronize_rcu();
        return ret;
 }