neigh: Convert garbage collection from softirq to workqueue
authorEric Dumazet <eric.dumazet@gmail.com>
Thu, 30 Jul 2009 03:15:07 +0000 (03:15 +0000)
committerDavid S. Miller <davem@davemloft.net>
Mon, 3 Aug 2009 01:35:16 +0000 (18:35 -0700)
Current neigh_periodic_timer() function is fired by timer IRQ, and
scans one hash bucket each round (very litle work in fact)

As we are supposed to scan whole hash table in 15 seconds, this means
neigh_periodic_timer() can be fired very often. (depending on the number
of concurrent hash entries we stored in this table)

Converting this to a workqueue permits scanning whole table, minimizing
icache pollution, and firing this work every 15 seconds, independantly
of hash table size.

This 15 seconds delay is not a hard number, as work is a deferrable one.

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

index d8d790e56d3daa2d7b1751220ca1fa4314a57346..18b69b6cecaf6e91d161d67057e1ae0fadfef1ef 100644 (file)
@@ -24,6 +24,7 @@
 
 #include <linux/err.h>
 #include <linux/sysctl.h>
+#include <linux/workqueue.h>
 #include <net/rtnetlink.h>
 
 /*
@@ -167,7 +168,7 @@ struct neigh_table
        int                     gc_thresh2;
        int                     gc_thresh3;
        unsigned long           last_flush;
-       struct timer_list       gc_timer;
+       struct delayed_work     gc_work;
        struct timer_list       proxy_timer;
        struct sk_buff_head     proxy_queue;
        atomic_t                entries;
@@ -178,7 +179,6 @@ struct neigh_table
        struct neighbour        **hash_buckets;
        unsigned int            hash_mask;
        __u32                   hash_rnd;
-       unsigned int            hash_chain_gc;
        struct pneigh_entry     **phash_buckets;
 };
 
index c6f9ad8e4c7a2e80d6e8fa34056907fad871ab26..e587e6819698cbe7485fb336cb70177e3bf98ca4 100644 (file)
@@ -692,75 +692,74 @@ static void neigh_connect(struct neighbour *neigh)
                hh->hh_output = neigh->ops->hh_output;
 }
 
-static void neigh_periodic_timer(unsigned long arg)
+static void neigh_periodic_work(struct work_struct *work)
 {
-       struct neigh_table *tbl = (struct neigh_table *)arg;
+       struct neigh_table *tbl = container_of(work, struct neigh_table, gc_work.work);
        struct neighbour *n, **np;
-       unsigned long expire, now = jiffies;
+       unsigned int i;
 
        NEIGH_CACHE_STAT_INC(tbl, periodic_gc_runs);
 
-       write_lock(&tbl->lock);
+       write_lock_bh(&tbl->lock);
 
        /*
         *      periodically recompute ReachableTime from random function
         */
 
-       if (time_after(now, tbl->last_rand + 300 * HZ)) {
+       if (time_after(jiffies, tbl->last_rand + 300 * HZ)) {
                struct neigh_parms *p;
-               tbl->last_rand = now;
+               tbl->last_rand = jiffies;
                for (p = &tbl->parms; p; p = p->next)
                        p->reachable_time =
                                neigh_rand_reach_time(p->base_reachable_time);
        }
 
-       np = &tbl->hash_buckets[tbl->hash_chain_gc];
-       tbl->hash_chain_gc = ((tbl->hash_chain_gc + 1) & tbl->hash_mask);
+       for (i = 0 ; i <= tbl->hash_mask; i++) {
+               np = &tbl->hash_buckets[i];
 
-       while ((n = *np) != NULL) {
-               unsigned int state;
+               while ((n = *np) != NULL) {
+                       unsigned int state;
 
-               write_lock(&n->lock);
+                       write_lock(&n->lock);
 
-               state = n->nud_state;
-               if (state & (NUD_PERMANENT | NUD_IN_TIMER)) {
-                       write_unlock(&n->lock);
-                       goto next_elt;
-               }
+                       state = n->nud_state;
+                       if (state & (NUD_PERMANENT | NUD_IN_TIMER)) {
+                               write_unlock(&n->lock);
+                               goto next_elt;
+                       }
 
-               if (time_before(n->used, n->confirmed))
-                       n->used = n->confirmed;
+                       if (time_before(n->used, n->confirmed))
+                               n->used = n->confirmed;
 
-               if (atomic_read(&n->refcnt) == 1 &&
-                   (state == NUD_FAILED ||
-                    time_after(now, n->used + n->parms->gc_staletime))) {
-                       *np = n->next;
-                       n->dead = 1;
+                       if (atomic_read(&n->refcnt) == 1 &&
+                           (state == NUD_FAILED ||
+                            time_after(jiffies, n->used + n->parms->gc_staletime))) {
+                               *np = n->next;
+                               n->dead = 1;
+                               write_unlock(&n->lock);
+                               neigh_cleanup_and_release(n);
+                               continue;
+                       }
                        write_unlock(&n->lock);
-                       neigh_cleanup_and_release(n);
-                       continue;
-               }
-               write_unlock(&n->lock);
 
 next_elt:
-               np = &n->next;
+                       np = &n->next;
+               }
+               /*
+                * It's fine to release lock here, even if hash table
+                * grows while we are preempted.
+                */
+               write_unlock_bh(&tbl->lock);
+               cond_resched();
+               write_lock_bh(&tbl->lock);
        }
-
        /* Cycle through all hash buckets every base_reachable_time/2 ticks.
         * ARP entry timeouts range from 1/2 base_reachable_time to 3/2
         * base_reachable_time.
         */
-       expire = tbl->parms.base_reachable_time >> 1;
-       expire /= (tbl->hash_mask + 1);
-       if (!expire)
-               expire = 1;
-
-       if (expire>HZ)
-               mod_timer(&tbl->gc_timer, round_jiffies(now + expire));
-       else
-               mod_timer(&tbl->gc_timer, now + expire);
-
-       write_unlock(&tbl->lock);
+       schedule_delayed_work(&tbl->gc_work,
+                             tbl->parms.base_reachable_time >> 1);
+       write_unlock_bh(&tbl->lock);
 }
 
 static __inline__ int neigh_max_probes(struct neighbour *n)
@@ -1442,10 +1441,8 @@ void neigh_table_init_no_netlink(struct neigh_table *tbl)
        get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
 
        rwlock_init(&tbl->lock);
-       setup_timer(&tbl->gc_timer, neigh_periodic_timer, (unsigned long)tbl);
-       tbl->gc_timer.expires  = now + 1;
-       add_timer(&tbl->gc_timer);
-
+       INIT_DELAYED_WORK_DEFERRABLE(&tbl->gc_work, neigh_periodic_work);
+       schedule_delayed_work(&tbl->gc_work, tbl->parms.reachable_time);
        setup_timer(&tbl->proxy_timer, neigh_proxy_process, (unsigned long)tbl);
        skb_queue_head_init_class(&tbl->proxy_queue,
                        &neigh_table_proxy_queue_class);
@@ -1482,7 +1479,8 @@ int neigh_table_clear(struct neigh_table *tbl)
        struct neigh_table **tp;
 
        /* It is not clean... Fix it to unload IPv6 module safely */
-       del_timer_sync(&tbl->gc_timer);
+       cancel_delayed_work(&tbl->gc_work);
+       flush_scheduled_work();
        del_timer_sync(&tbl->proxy_timer);
        pneigh_queue_purge(&tbl->proxy_queue);
        neigh_ifdown(tbl, NULL);
@@ -1752,7 +1750,6 @@ static int neightbl_fill_info(struct sk_buff *skb, struct neigh_table *tbl,
                        .ndtc_last_rand         = jiffies_to_msecs(rand_delta),
                        .ndtc_hash_rnd          = tbl->hash_rnd,
                        .ndtc_hash_mask         = tbl->hash_mask,
-                       .ndtc_hash_chain_gc     = tbl->hash_chain_gc,
                        .ndtc_proxy_qlen        = tbl->proxy_queue.qlen,
                };