flow: delayed deletion of flow cache entries
Timo Teräs [Wed, 7 Apr 2010 00:30:07 +0000 (00:30 +0000)]
Speed up lookups by freeing flow cache entries later. After
virtualizing flow cache entry operations, the flow cache may now
end up calling policy or bundle destructor which can be slowish.

As gc_list is more effective with double linked list, the flow cache
is converted to use common hlist and list macroes where appropriate.

Signed-off-by: Timo Teras <timo.teras@iki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/core/flow.c

index 521df52..1619006 100644 (file)
 #include <linux/security.h>
 
 struct flow_cache_entry {
-       struct flow_cache_entry         *next;
+       union {
+               struct hlist_node       hlist;
+               struct list_head        gc_list;
+       } u;
        u16                             family;
        u8                              dir;
        u32                             genid;
@@ -35,7 +38,7 @@ struct flow_cache_entry {
 };
 
 struct flow_cache_percpu {
-       struct flow_cache_entry         **hash_table;
+       struct hlist_head               *hash_table;
        int                             hash_count;
        u32                             hash_rnd;
        int                             hash_rnd_recalc;
@@ -62,6 +65,9 @@ atomic_t flow_cache_genid = ATOMIC_INIT(0);
 static struct flow_cache flow_cache_global;
 static struct kmem_cache *flow_cachep;
 
+static DEFINE_SPINLOCK(flow_cache_gc_lock);
+static LIST_HEAD(flow_cache_gc_list);
+
 #define flow_cache_hash_size(cache)    (1 << (cache)->hash_shift)
 #define FLOW_HASH_RND_PERIOD           (10 * 60 * HZ)
 
@@ -86,38 +92,66 @@ static int flow_entry_valid(struct flow_cache_entry *fle)
        return 1;
 }
 
-static void flow_entry_kill(struct flow_cache *fc,
-                           struct flow_cache_percpu *fcp,
-                           struct flow_cache_entry *fle)
+static void flow_entry_kill(struct flow_cache_entry *fle)
 {
        if (fle->object)
                fle->object->ops->delete(fle->object);
        kmem_cache_free(flow_cachep, fle);
-       fcp->hash_count--;
+}
+
+static void flow_cache_gc_task(struct work_struct *work)
+{
+       struct list_head gc_list;
+       struct flow_cache_entry *fce, *n;
+
+       INIT_LIST_HEAD(&gc_list);
+       spin_lock_bh(&flow_cache_gc_lock);
+       list_splice_tail_init(&flow_cache_gc_list, &gc_list);
+       spin_unlock_bh(&flow_cache_gc_lock);
+
+       list_for_each_entry_safe(fce, n, &gc_list, u.gc_list)
+               flow_entry_kill(fce);
+}
+static DECLARE_WORK(flow_cache_gc_work, flow_cache_gc_task);
+
+static void flow_cache_queue_garbage(struct flow_cache_percpu *fcp,
+                                    int deleted, struct list_head *gc_list)
+{
+       if (deleted) {
+               fcp->hash_count -= deleted;
+               spin_lock_bh(&flow_cache_gc_lock);
+               list_splice_tail(gc_list, &flow_cache_gc_list);
+               spin_unlock_bh(&flow_cache_gc_lock);
+               schedule_work(&flow_cache_gc_work);
+       }
 }
 
 static void __flow_cache_shrink(struct flow_cache *fc,
                                struct flow_cache_percpu *fcp,
                                int shrink_to)
 {
-       struct flow_cache_entry *fle, **flp;
-       int i;
+       struct flow_cache_entry *fle;
+       struct hlist_node *entry, *tmp;
+       LIST_HEAD(gc_list);
+       int i, deleted = 0;
 
        for (i = 0; i < flow_cache_hash_size(fc); i++) {
                int saved = 0;
 
-               flp = &fcp->hash_table[i];
-               while ((fle = *flp) != NULL) {
+               hlist_for_each_entry_safe(fle, entry, tmp,
+                                         &fcp->hash_table[i], u.hlist) {
                        if (saved < shrink_to &&
                            flow_entry_valid(fle)) {
                                saved++;
-                               flp = &fle->next;
                        } else {
-                               *flp = fle->next;
-                               flow_entry_kill(fc, fcp, fle);
+                               deleted++;
+                               hlist_del(&fle->u.hlist);
+                               list_add_tail(&fle->u.gc_list, &gc_list);
                        }
                }
        }
+
+       flow_cache_queue_garbage(fcp, deleted, &gc_list);
 }
 
 static void flow_cache_shrink(struct flow_cache *fc,
@@ -182,7 +216,8 @@ flow_cache_lookup(struct net *net, struct flowi *key, u16 family, u8 dir,
 {
        struct flow_cache *fc = &flow_cache_global;
        struct flow_cache_percpu *fcp;
-       struct flow_cache_entry *fle, **head;
+       struct flow_cache_entry *fle, *tfle;
+       struct hlist_node *entry;
        struct flow_cache_object *flo;
        unsigned int hash;
 
@@ -200,12 +235,13 @@ flow_cache_lookup(struct net *net, struct flowi *key, u16 family, u8 dir,
                flow_new_hash_rnd(fc, fcp);
 
        hash = flow_hash_code(fc, fcp, key);
-       head = &fcp->hash_table[hash];
-       for (fle = *head; fle; fle = fle->next) {
-               if (fle->family == family &&
-                   fle->dir == dir &&
-                   flow_key_compare(key, &fle->key) == 0)
+       hlist_for_each_entry(tfle, entry, &fcp->hash_table[hash], u.hlist) {
+               if (tfle->family == family &&
+                   tfle->dir == dir &&
+                   flow_key_compare(key, &tfle->key) == 0) {
+                       fle = tfle;
                        break;
+               }
        }
 
        if (unlikely(!fle)) {
@@ -214,12 +250,11 @@ flow_cache_lookup(struct net *net, struct flowi *key, u16 family, u8 dir,
 
                fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
                if (fle) {
-                       fle->next = *head;
-                       *head = fle;
                        fle->family = family;
                        fle->dir = dir;
                        memcpy(&fle->key, key, sizeof(*key));
                        fle->object = NULL;
+                       hlist_add_head(&fle->u.hlist, &fcp->hash_table[hash]);
                        fcp->hash_count++;
                }
        } else if (likely(fle->genid == atomic_read(&flow_cache_genid))) {
@@ -262,23 +297,26 @@ static void flow_cache_flush_tasklet(unsigned long data)
        struct flow_flush_info *info = (void *)data;
        struct flow_cache *fc = info->cache;
        struct flow_cache_percpu *fcp;
-       int i;
+       struct flow_cache_entry *fle;
+       struct hlist_node *entry, *tmp;
+       LIST_HEAD(gc_list);
+       int i, deleted = 0;
 
        fcp = per_cpu_ptr(fc->percpu, smp_processor_id());
        for (i = 0; i < flow_cache_hash_size(fc); i++) {
-               struct flow_cache_entry *fle;
-
-               fle = fcp->hash_table[i];
-               for (; fle; fle = fle->next) {
+               hlist_for_each_entry_safe(fle, entry, tmp,
+                                         &fcp->hash_table[i], u.hlist) {
                        if (flow_entry_valid(fle))
                                continue;
 
-                       if (fle->object)
-                               fle->object->ops->delete(fle->object);
-                       fle->object = NULL;
+                       deleted++;
+                       hlist_del(&fle->u.hlist);
+                       list_add_tail(&fle->u.gc_list, &gc_list);
                }
        }
 
+       flow_cache_queue_garbage(fcp, deleted, &gc_list);
+
        if (atomic_dec_and_test(&info->cpuleft))
                complete(&info->completion);
 }
@@ -320,7 +358,7 @@ void flow_cache_flush(void)
 static void __init flow_cache_cpu_prepare(struct flow_cache *fc,
                                          struct flow_cache_percpu *fcp)
 {
-       fcp->hash_table = (struct flow_cache_entry **)
+       fcp->hash_table = (struct hlist_head *)
                __get_free_pages(GFP_KERNEL|__GFP_ZERO, fc->order);
        if (!fcp->hash_table)
                panic("NET: failed to allocate flow cache order %lu\n", fc->order);
@@ -354,7 +392,7 @@ static int flow_cache_init(struct flow_cache *fc)
 
        for (order = 0;
             (PAGE_SIZE << order) <
-                    (sizeof(struct flow_cache_entry *)*flow_cache_hash_size(fc));
+                    (sizeof(struct hlist_head)*flow_cache_hash_size(fc));
             order++)
                /* NOTHING */;
        fc->order = order;