[PATCH] IPV4 : Move ip route cache flush (secret_rebuild) from softirq to workqueue
authorEric Dumazet <dada1@cosmosbay.com>
Tue, 20 Nov 2007 06:43:37 +0000 (22:43 -0800)
committerDavid S. Miller <davem@davemloft.net>
Mon, 28 Jan 2008 22:54:33 +0000 (14:54 -0800)
Every 600 seconds (ip_rt_secret_interval), a softirq flush of the
whole ip route cache is triggered. On loaded machines, this can starve
softirq for many seconds and can eventually crash.

This patch moves this flush to a workqueue context, using the worker
we intoduced in commit 39c90ece7565f5c47110c2fa77409d7a9478bd5b (IPV4:
Convert rt_check_expire() from softirq processing to workqueue.)

Also, immediate flushes (echo 0 >/proc/sys/net/ipv4/route/flush) are
using rt_do_flush() helper function, wich take attention to
rescheduling.

Next step will be to handle delayed flushes
("echo -1 >/proc/sys/net/ipv4/route/flush" or "ip route flush cache")

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/route.c

index e4b6fb4b1f4f983953c354eff95ec8af57241f56..fcae074b7ae4f83e87d8e41d6e48b356cea58c27 100644 (file)
@@ -133,13 +133,14 @@ static int ip_rt_mtu_expires              = 10 * 60 * HZ;
 static int ip_rt_min_pmtu              = 512 + 20 + 20;
 static int ip_rt_min_advmss            = 256;
 static int ip_rt_secret_interval       = 10 * 60 * HZ;
+static int ip_rt_flush_expected;
 static unsigned long rt_deadline;
 
 #define RTprint(a...)  printk(KERN_DEBUG a)
 
 static struct timer_list rt_flush_timer;
-static void rt_check_expire(struct work_struct *work);
-static DECLARE_DELAYED_WORK(expires_work, rt_check_expire);
+static void rt_worker_func(struct work_struct *work);
+static DECLARE_DELAYED_WORK(expires_work, rt_worker_func);
 static struct timer_list rt_secret_timer;
 
 /*
@@ -561,7 +562,36 @@ static inline int compare_keys(struct flowi *fl1, struct flowi *fl2)
                (fl1->iif ^ fl2->iif)) == 0;
 }
 
-static void rt_check_expire(struct work_struct *work)
+/*
+ * Perform a full scan of hash table and free all entries.
+ * Can be called by a softirq or a process.
+ * In the later case, we want to be reschedule if necessary
+ */
+static void rt_do_flush(int process_context)
+{
+       unsigned int i;
+       struct rtable *rth, *next;
+
+       for (i = 0; i <= rt_hash_mask; i++) {
+               if (process_context && need_resched())
+                       cond_resched();
+               rth = rt_hash_table[i].chain;
+               if (!rth)
+                       continue;
+
+               spin_lock_bh(rt_hash_lock_addr(i));
+               rth = rt_hash_table[i].chain;
+               rt_hash_table[i].chain = NULL;
+               spin_unlock_bh(rt_hash_lock_addr(i));
+
+               for (; rth; rth = next) {
+                       next = rth->u.dst.rt_next;
+                       rt_free(rth);
+               }
+       }
+}
+
+static void rt_check_expire(void)
 {
        static unsigned int rover;
        unsigned int i = rover, goal;
@@ -607,33 +637,33 @@ static void rt_check_expire(struct work_struct *work)
                spin_unlock_bh(rt_hash_lock_addr(i));
        }
        rover = i;
+}
+
+/*
+ * rt_worker_func() is run in process context.
+ * If a whole flush was scheduled, it is done.
+ * Else, we call rt_check_expire() to scan part of the hash table
+ */
+static void rt_worker_func(struct work_struct *work)
+{
+       if (ip_rt_flush_expected) {
+               ip_rt_flush_expected = 0;
+               rt_do_flush(1);
+       } else
+               rt_check_expire();
        schedule_delayed_work(&expires_work, ip_rt_gc_interval);
 }
 
 /* This can run from both BH and non-BH contexts, the latter
  * in the case of a forced flush event.
  */
-static void rt_run_flush(unsigned long dummy)
+static void rt_run_flush(unsigned long process_context)
 {
-       int i;
-       struct rtable *rth, *next;
-
        rt_deadline = 0;
 
        get_random_bytes(&rt_hash_rnd, 4);
 
-       for (i = rt_hash_mask; i >= 0; i--) {
-               spin_lock_bh(rt_hash_lock_addr(i));
-               rth = rt_hash_table[i].chain;
-               if (rth)
-                       rt_hash_table[i].chain = NULL;
-               spin_unlock_bh(rt_hash_lock_addr(i));
-
-               for (; rth; rth = next) {
-                       next = rth->u.dst.rt_next;
-                       rt_free(rth);
-               }
-       }
+       rt_do_flush(process_context);
 }
 
 static DEFINE_SPINLOCK(rt_flush_lock);
@@ -667,7 +697,7 @@ void rt_cache_flush(int delay)
 
        if (delay <= 0) {
                spin_unlock_bh(&rt_flush_lock);
-               rt_run_flush(0);
+               rt_run_flush(user_mode);
                return;
        }
 
@@ -678,12 +708,17 @@ void rt_cache_flush(int delay)
        spin_unlock_bh(&rt_flush_lock);
 }
 
+/*
+ * We change rt_hash_rnd and ask next rt_worker_func() invocation
+ * to perform a flush in process context
+ */
 static void rt_secret_rebuild(unsigned long dummy)
 {
-       unsigned long now = jiffies;
-
-       rt_cache_flush(0);
-       mod_timer(&rt_secret_timer, now + ip_rt_secret_interval);
+       get_random_bytes(&rt_hash_rnd, 4);
+       ip_rt_flush_expected = 1;
+       cancel_delayed_work(&expires_work);
+       schedule_delayed_work(&expires_work, HZ/10);
+       mod_timer(&rt_secret_timer, jiffies + ip_rt_secret_interval);
 }
 
 /*