]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commitdiff
fib_hash: RCU conversion phase 2
authorEric Dumazet <eric.dumazet@gmail.com>
Thu, 14 Oct 2010 20:56:39 +0000 (20:56 +0000)
committerDavid S. Miller <davem@davemloft.net>
Sun, 17 Oct 2010 20:53:16 +0000 (13:53 -0700)
Get rid of fib_hash_lock rwlock.

The fn_zone hash table resize is the noticeable part of this patch.

I added a seqlock per fn_zone, so that readers can restart their lookup
in the (very rare) case a writer expanded the hash table.

Add rcu heads in fib_alias and fib_node, use call_rcu() to defer their
freeing, and use appropriate _rcu list manipulations.

Stress test (160.000.000 udp frames sent, IP route cache disabled to
mimic DDOS attack, FIB_HASH)

Before:
real 0m41.191s
user 0m13.137s
sys 8m55.241s

After:
real 0m38.091s
user 0m13.189s
sys 7m53.018s

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/fib_hash.c
net/ipv4/fib_lookup.h

index 04f05a96b75b16cc317aefbabaafae3769f31f8a..4f1aafd3ba89f92a65d01488a37591c35b0f669c 100644 (file)
@@ -58,7 +58,8 @@ struct fib_node {
 
 struct fn_zone {
        struct fn_zone __rcu    *fz_next;       /* Next not empty zone  */
-       struct hlist_head       *fz_hash;       /* Hash table pointer   */
+       struct hlist_head __rcu *fz_hash;       /* Hash table pointer   */
+       seqlock_t               fz_lock;
        u32                     fz_hashmask;    /* (fz_divisor - 1)     */
 
        u8                      fz_order;       /* Zone order (0..32)   */
@@ -92,7 +93,6 @@ static inline __be32 fz_key(__be32 dst, struct fn_zone *fz)
        return dst & FZ_MASK(fz);
 }
 
-static DEFINE_RWLOCK(fib_hash_lock);
 static unsigned int fib_hash_genid;
 
 #define FZ_MAX_DIVISOR ((PAGE_SIZE<<MAX_ORDER) / sizeof(struct hlist_head))
@@ -101,12 +101,11 @@ static struct hlist_head *fz_hash_alloc(int divisor)
 {
        unsigned long size = divisor * sizeof(struct hlist_head);
 
-       if (size <= PAGE_SIZE) {
+       if (size <= PAGE_SIZE)
                return kzalloc(size, GFP_KERNEL);
-       } else {
-               return (struct hlist_head *)
-                       __get_free_pages(GFP_KERNEL | __GFP_ZERO, get_order(size));
-       }
+
+       return (struct hlist_head *)
+               __get_free_pages(GFP_KERNEL | __GFP_ZERO, get_order(size));
 }
 
 /* The fib hash lock must be held when this is called. */
@@ -121,12 +120,12 @@ static inline void fn_rebuild_zone(struct fn_zone *fz,
                struct fib_node *f;
 
                hlist_for_each_entry_safe(f, node, n, &old_ht[i], fn_hash) {
-                       struct hlist_head *new_head;
+                       struct hlist_head __rcu *new_head;
 
-                       hlist_del(&f->fn_hash);
+                       hlist_del_rcu(&f->fn_hash);
 
                        new_head = &fz->fz_hash[fn_hash(f->fn_key, fz)];
-                       hlist_add_head(&f->fn_hash, new_head);
+                       hlist_add_head_rcu(&f->fn_hash, new_head);
                }
        }
 }
@@ -175,32 +174,55 @@ static void fn_rehash_zone(struct fn_zone *fz)
        ht = fz_hash_alloc(new_divisor);
 
        if (ht) {
-               write_lock_bh(&fib_hash_lock);
+               struct fn_zone nfz;
+
+               memcpy(&nfz, fz, sizeof(nfz));
+
+               write_seqlock_bh(&fz->fz_lock);
                old_ht = fz->fz_hash;
-               fz->fz_hash = ht;
+               nfz.fz_hash = ht;
+               nfz.fz_hashmask = new_hashmask;
+               nfz.fz_divisor = new_divisor;
+               fn_rebuild_zone(&nfz, old_ht, old_divisor);
+               fib_hash_genid++;
+               rcu_assign_pointer(fz->fz_hash, ht);
                fz->fz_hashmask = new_hashmask;
                fz->fz_divisor = new_divisor;
-               fn_rebuild_zone(fz, old_ht, old_divisor);
-               fib_hash_genid++;
-               write_unlock_bh(&fib_hash_lock);
+               write_sequnlock_bh(&fz->fz_lock);
 
-               if (old_ht != fz->fz_embedded_hash)
+               if (old_ht != fz->fz_embedded_hash) {
+                       synchronize_rcu();
                        fz_hash_free(old_ht, old_divisor);
+               }
        }
 }
 
-static inline void fn_free_node(struct fib_node * f)
+static void fn_free_node_rcu(struct rcu_head *head)
 {
+       struct fib_node *f = container_of(head, struct fib_node, fn_embedded_alias.rcu);
+
        kmem_cache_free(fn_hash_kmem, f);
 }
 
+static inline void fn_free_node(struct fib_node *f)
+{
+       call_rcu(&f->fn_embedded_alias.rcu, fn_free_node_rcu);
+}
+
+static void fn_free_alias_rcu(struct rcu_head *head)
+{
+       struct fib_alias *fa = container_of(head, struct fib_alias, rcu);
+
+       kmem_cache_free(fn_alias_kmem, fa);
+}
+
 static inline void fn_free_alias(struct fib_alias *fa, struct fib_node *f)
 {
        fib_release_info(fa->fa_info);
        if (fa == &f->fn_embedded_alias)
                fa->fa_info = NULL;
        else
-               kmem_cache_free(fn_alias_kmem, fa);
+               call_rcu(&fa->rcu, fn_free_alias_rcu);
 }
 
 static struct fn_zone *
@@ -211,6 +233,7 @@ fn_new_zone(struct fn_hash *table, int z)
        if (!fz)
                return NULL;
 
+       seqlock_init(&fz->fz_lock);
        fz->fz_divisor = z ? EMBEDDED_HASH_SIZE : 1;
        fz->fz_hashmask = fz->fz_divisor - 1;
        fz->fz_hash = fz->fz_embedded_hash;
@@ -246,30 +269,34 @@ int fib_table_lookup(struct fib_table *tb,
        struct fn_hash *t = (struct fn_hash *)tb->tb_data;
 
        rcu_read_lock();
-       read_lock(&fib_hash_lock);
        for (fz = rcu_dereference(t->fn_zone_list);
             fz != NULL;
             fz = rcu_dereference(fz->fz_next)) {
-               struct hlist_head *head;
+               struct hlist_head __rcu *head;
                struct hlist_node *node;
                struct fib_node *f;
-               __be32 k = fz_key(flp->fl4_dst, fz);
+               __be32 k;
+               unsigned int seq;
 
-               head = &fz->fz_hash[fn_hash(k, fz)];
-               hlist_for_each_entry(f, node, head, fn_hash) {
-                       if (f->fn_key != k)
-                               continue;
+               do {
+                       seq = read_seqbegin(&fz->fz_lock);
+                       k = fz_key(flp->fl4_dst, fz);
+
+                       head = &fz->fz_hash[fn_hash(k, fz)];
+                       hlist_for_each_entry_rcu(f, node, head, fn_hash) {
+                               if (f->fn_key != k)
+                                       continue;
 
-                       err = fib_semantic_match(&f->fn_alias,
+                               err = fib_semantic_match(&f->fn_alias,
                                                 flp, res,
                                                 fz->fz_order, fib_flags);
-                       if (err <= 0)
-                               goto out;
-               }
+                               if (err <= 0)
+                                       goto out;
+                       }
+               } while (read_seqretry(&fz->fz_lock, seq));
        }
        err = 1;
 out:
-       read_unlock(&fib_hash_lock);
        rcu_read_unlock();
        return err;
 }
@@ -292,11 +319,11 @@ void fib_table_select_default(struct fib_table *tb,
        last_resort = NULL;
        order = -1;
 
-       read_lock(&fib_hash_lock);
-       hlist_for_each_entry(f, node, &fz->fz_hash[0], fn_hash) {
+       rcu_read_lock();
+       hlist_for_each_entry_rcu(f, node, &fz->fz_hash[0], fn_hash) {
                struct fib_alias *fa;
 
-               list_for_each_entry(fa, &f->fn_alias, fa_list) {
+               list_for_each_entry_rcu(fa, &f->fn_alias, fa_list) {
                        struct fib_info *next_fi = fa->fa_info;
 
                        if (fa->fa_scope != res->scope ||
@@ -340,7 +367,7 @@ void fib_table_select_default(struct fib_table *tb,
                fib_result_assign(res, last_resort);
        tb->tb_default = last_idx;
 out:
-       read_unlock(&fib_hash_lock);
+       rcu_read_unlock();
 }
 
 /* Insert node F to FZ. */
@@ -348,7 +375,7 @@ static inline void fib_insert_node(struct fn_zone *fz, struct fib_node *f)
 {
        struct hlist_head *head = &fz->fz_hash[fn_hash(f->fn_key, fz)];
 
-       hlist_add_head(&f->fn_hash, head);
+       hlist_add_head_rcu(&f->fn_hash, head);
 }
 
 /* Return the node in FZ matching KEY. */
@@ -358,7 +385,7 @@ static struct fib_node *fib_find_node(struct fn_zone *fz, __be32 key)
        struct hlist_node *node;
        struct fib_node *f;
 
-       hlist_for_each_entry(f, node, head, fn_hash) {
+       hlist_for_each_entry_rcu(f, node, head, fn_hash) {
                if (f->fn_key == key)
                        return f;
        }
@@ -366,6 +393,16 @@ static struct fib_node *fib_find_node(struct fn_zone *fz, __be32 key)
        return NULL;
 }
 
+
+static struct fib_alias *fib_fast_alloc(struct fib_node *f)
+{
+       struct fib_alias *fa = &f->fn_embedded_alias;
+
+       if (fa->fa_info != NULL)
+               fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
+       return fa;
+}
+
 /* Caller must hold RTNL. */
 int fib_table_insert(struct fib_table *tb, struct fib_config *cfg)
 {
@@ -451,7 +488,6 @@ int fib_table_insert(struct fib_table *tb, struct fib_config *cfg)
                }
 
                if (cfg->fc_nlflags & NLM_F_REPLACE) {
-                       struct fib_info *fi_drop;
                        u8 state;
 
                        fa = fa_first;
@@ -460,21 +496,25 @@ int fib_table_insert(struct fib_table *tb, struct fib_config *cfg)
                                        err = 0;
                                goto out;
                        }
-                       write_lock_bh(&fib_hash_lock);
-                       fi_drop = fa->fa_info;
-                       fa->fa_info = fi;
-                       fa->fa_type = cfg->fc_type;
-                       fa->fa_scope = cfg->fc_scope;
+                       err = -ENOBUFS;
+                       new_fa = fib_fast_alloc(f);
+                       if (new_fa == NULL)
+                               goto out;
+
+                       new_fa->fa_tos = fa->fa_tos;
+                       new_fa->fa_info = fi;
+                       new_fa->fa_type = cfg->fc_type;
+                       new_fa->fa_scope = cfg->fc_scope;
                        state = fa->fa_state;
-                       fa->fa_state &= ~FA_S_ACCESSED;
+                       new_fa->fa_state = state & ~FA_S_ACCESSED;
                        fib_hash_genid++;
-                       write_unlock_bh(&fib_hash_lock);
+                       list_replace_rcu(&fa->fa_list, &new_fa->fa_list);
 
-                       fib_release_info(fi_drop);
+                       fn_free_alias(fa, f);
                        if (state & FA_S_ACCESSED)
                                rt_cache_flush(cfg->fc_nlinfo.nl_net, -1);
-                       rtmsg_fib(RTM_NEWROUTE, key, fa, cfg->fc_dst_len, tb->tb_id,
-                                 &cfg->fc_nlinfo, NLM_F_REPLACE);
+                       rtmsg_fib(RTM_NEWROUTE, key, new_fa, cfg->fc_dst_len,
+                                 tb->tb_id, &cfg->fc_nlinfo, NLM_F_REPLACE);
                        return 0;
                }
 
@@ -506,12 +546,10 @@ int fib_table_insert(struct fib_table *tb, struct fib_config *cfg)
                f = new_f;
        }
 
-       new_fa = &f->fn_embedded_alias;
-       if (new_fa->fa_info != NULL) {
-               new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
-               if (new_fa == NULL)
-                       goto out;
-       }
+       new_fa = fib_fast_alloc(f);
+       if (new_fa == NULL)
+               goto out;
+
        new_fa->fa_info = fi;
        new_fa->fa_tos = tos;
        new_fa->fa_type = cfg->fc_type;
@@ -522,13 +560,11 @@ int fib_table_insert(struct fib_table *tb, struct fib_config *cfg)
         * Insert new entry to the list.
         */
 
-       write_lock_bh(&fib_hash_lock);
        if (new_f)
                fib_insert_node(fz, new_f);
-       list_add_tail(&new_fa->fa_list,
+       list_add_tail_rcu(&new_fa->fa_list,
                 (fa ? &fa->fa_list : &f->fn_alias));
        fib_hash_genid++;
-       write_unlock_bh(&fib_hash_lock);
 
        if (new_f)
                fz->fz_nent++;
@@ -603,14 +639,12 @@ int fib_table_delete(struct fib_table *tb, struct fib_config *cfg)
                          tb->tb_id, &cfg->fc_nlinfo, 0);
 
                kill_fn = 0;
-               write_lock_bh(&fib_hash_lock);
-               list_del(&fa->fa_list);
+               list_del_rcu(&fa->fa_list);
                if (list_empty(&f->fn_alias)) {
-                       hlist_del(&f->fn_hash);
+                       hlist_del_rcu(&f->fn_hash);
                        kill_fn = 1;
                }
                fib_hash_genid++;
-               write_unlock_bh(&fib_hash_lock);
 
                if (fa->fa_state & FA_S_ACCESSED)
                        rt_cache_flush(cfg->fc_nlinfo.nl_net, -1);
@@ -641,14 +675,12 @@ static int fn_flush_list(struct fn_zone *fz, int idx)
                        struct fib_info *fi = fa->fa_info;
 
                        if (fi && (fi->fib_flags&RTNH_F_DEAD)) {
-                               write_lock_bh(&fib_hash_lock);
-                               list_del(&fa->fa_list);
+                               list_del_rcu(&fa->fa_list);
                                if (list_empty(&f->fn_alias)) {
-                                       hlist_del(&f->fn_hash);
+                                       hlist_del_rcu(&f->fn_hash);
                                        kill_f = 1;
                                }
                                fib_hash_genid++;
-                               write_unlock_bh(&fib_hash_lock);
 
                                fn_free_alias(fa, f);
                                found++;
@@ -693,10 +725,10 @@ fn_hash_dump_bucket(struct sk_buff *skb, struct netlink_callback *cb,
 
        s_i = cb->args[4];
        i = 0;
-       hlist_for_each_entry(f, node, head, fn_hash) {
+       hlist_for_each_entry_rcu(f, node, head, fn_hash) {
                struct fib_alias *fa;
 
-               list_for_each_entry(fa, &f->fn_alias, fa_list) {
+               list_for_each_entry_rcu(fa, &f->fn_alias, fa_list) {
                        if (i < s_i)
                                goto next;
 
@@ -714,7 +746,7 @@ fn_hash_dump_bucket(struct sk_buff *skb, struct netlink_callback *cb,
                                cb->args[4] = i;
                                return -1;
                        }
-               next:
+next:
                        i++;
                }
        }
@@ -755,7 +787,6 @@ int fib_table_dump(struct fib_table *tb, struct sk_buff *skb,
 
        s_m = cb->args[2];
        rcu_read_lock();
-       read_lock(&fib_hash_lock);
        for (fz = rcu_dereference(table->fn_zone_list);
             fz != NULL;
             fz = rcu_dereference(fz->fz_next), m++) {
@@ -763,14 +794,12 @@ int fib_table_dump(struct fib_table *tb, struct sk_buff *skb,
                        continue;
                if (fn_hash_dump_zone(skb, cb, tb, fz) < 0) {
                        cb->args[2] = m;
-                       read_unlock(&fib_hash_lock);
                        rcu_read_unlock();
                        return -1;
                }
                memset(&cb->args[3], 0,
                       sizeof(cb->args) - 3*sizeof(cb->args[0]));
        }
-       read_unlock(&fib_hash_lock);
        rcu_read_unlock();
        cb->args[2] = m;
        return skb->len;
@@ -960,13 +989,11 @@ static struct fib_alias *fib_get_idx(struct seq_file *seq, loff_t pos)
 }
 
 static void *fib_seq_start(struct seq_file *seq, loff_t *pos)
-       __acquires(fib_hash_lock)
        __acquires(RCU)
 {
        void *v = NULL;
 
        rcu_read_lock();
-       read_lock(&fib_hash_lock);
        if (fib_get_table(seq_file_net(seq), RT_TABLE_MAIN))
                v = *pos ? fib_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
        return v;
@@ -979,17 +1006,16 @@ static void *fib_seq_next(struct seq_file *seq, void *v, loff_t *pos)
 }
 
 static void fib_seq_stop(struct seq_file *seq, void *v)
-       __releases(fib_hash_lock)
        __releases(RCU)
 {
-       read_unlock(&fib_hash_lock);
        rcu_read_unlock();
 }
 
 static unsigned fib_flag_trans(int type, __be32 mask, struct fib_info *fi)
 {
        static const unsigned type2flags[RTN_MAX + 1] = {
-               [7] = RTF_REJECT, [8] = RTF_REJECT,
+               [7] = RTF_REJECT,
+               [8] = RTF_REJECT,
        };
        unsigned flags = type2flags[type];
 
index b9c9a9f2aee54f50f14f7594f119f46b3631fe3e..5072d8effd5d21a75a6b63494f3630bfb48de921 100644 (file)
@@ -12,9 +12,7 @@ struct fib_alias {
        u8                      fa_type;
        u8                      fa_scope;
        u8                      fa_state;
-#ifdef CONFIG_IP_FIB_TRIE
        struct rcu_head         rcu;
-#endif
 };
 
 #define FA_S_ACCESSED  0x01