[NET_SCHED]: Fix fallout from dev->qdisc RCU change
authorPatrick McHardy <kaber@trash.net>
Wed, 27 Sep 2006 23:45:45 +0000 (16:45 -0700)
committerDavid S. Miller <davem@sunset.davemloft.net>
Fri, 29 Sep 2006 01:01:50 +0000 (18:01 -0700)
The move of qdisc destruction to a rcu callback broke locking in the
entire qdisc layer by invalidating previously valid assumptions about
the context in which changes to the qdisc tree occur.

The two assumptions were:

- since changes only happen in process context, read_lock doesn't need
  bottem half protection. Now invalid since destruction of inner qdiscs,
  classifiers, actions and estimators happens in the RCU callback unless
  they're manually deleted, resulting in dead-locks when read_lock in
  process context is interrupted by write_lock_bh in bottem half context.

- since changes only happen under the RTNL, no additional locking is
  necessary for data not used during packet processing (f.e. u32_list).
  Again, since destruction now happens in the RCU callback, this assumption
  is not valid anymore, causing races while using this data, which can
  result in corruption or use-after-free.

Instead of "fixing" this by disabling bottem halfs everywhere and adding
new locks/refcounting, this patch makes these assumptions valid again by
moving destruction back to process context. Since only the dev->qdisc
pointer is protected by RCU, but ->enqueue and the qdisc tree are still
protected by dev->qdisc_lock, destruction of the tree can be performed
immediately and only the final free needs to happen in the rcu callback
to make sure dev_queue_xmit doesn't access already freed memory.

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/dev.c
net/sched/cls_api.c
net/sched/sch_api.c
net/sched/sch_generic.c

index 14de297d024d1e8771180dde2e07aec624b90e1c..4d891beab13899cd96c74b8a6a88b5ef5d6aba43 100644 (file)
@@ -1480,14 +1480,16 @@ gso:
        if (q->enqueue) {
                /* Grab device queue */
                spin_lock(&dev->queue_lock);
+               q = dev->qdisc;
+               if (q->enqueue) {
+                       rc = q->enqueue(skb, q);
+                       qdisc_run(dev);
+                       spin_unlock(&dev->queue_lock);
 
-               rc = q->enqueue(skb, q);
-
-               qdisc_run(dev);
-
+                       rc = rc == NET_XMIT_BYPASS ? NET_XMIT_SUCCESS : rc;
+                       goto out;
+               }
                spin_unlock(&dev->queue_lock);
-               rc = rc == NET_XMIT_BYPASS ? NET_XMIT_SUCCESS : rc;
-               goto out;
        }
 
        /* The device has no queue. Common case for software devices:
index 7e14f14058e991393ca1a75aacc669a516e1906c..37a1840216474d77c2ad534903059bbeb1a840ea 100644 (file)
@@ -401,7 +401,7 @@ static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
        if ((dev = dev_get_by_index(tcm->tcm_ifindex)) == NULL)
                return skb->len;
 
-       read_lock_bh(&qdisc_tree_lock);
+       read_lock(&qdisc_tree_lock);
        if (!tcm->tcm_parent)
                q = dev->qdisc_sleeping;
        else
@@ -458,7 +458,7 @@ errout:
        if (cl)
                cops->put(q, cl);
 out:
-       read_unlock_bh(&qdisc_tree_lock);
+       read_unlock(&qdisc_tree_lock);
        dev_put(dev);
        return skb->len;
 }
index a19eff12cf78b9013d66f070153c869f8242b151..0b6489291140f0324c73e8a8ddb0aad716c03593 100644 (file)
@@ -195,14 +195,14 @@ struct Qdisc *qdisc_lookup(struct net_device *dev, u32 handle)
 {
        struct Qdisc *q;
 
-       read_lock_bh(&qdisc_tree_lock);
+       read_lock(&qdisc_tree_lock);
        list_for_each_entry(q, &dev->qdisc_list, list) {
                if (q->handle == handle) {
-                       read_unlock_bh(&qdisc_tree_lock);
+                       read_unlock(&qdisc_tree_lock);
                        return q;
                }
        }
-       read_unlock_bh(&qdisc_tree_lock);
+       read_unlock(&qdisc_tree_lock);
        return NULL;
 }
 
@@ -837,7 +837,7 @@ static int tc_dump_qdisc(struct sk_buff *skb, struct netlink_callback *cb)
                        continue;
                if (idx > s_idx)
                        s_q_idx = 0;
-               read_lock_bh(&qdisc_tree_lock);
+               read_lock(&qdisc_tree_lock);
                q_idx = 0;
                list_for_each_entry(q, &dev->qdisc_list, list) {
                        if (q_idx < s_q_idx) {
@@ -846,12 +846,12 @@ static int tc_dump_qdisc(struct sk_buff *skb, struct netlink_callback *cb)
                        }
                        if (tc_fill_qdisc(skb, q, q->parent, NETLINK_CB(cb->skb).pid,
                                          cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWQDISC) <= 0) {
-                               read_unlock_bh(&qdisc_tree_lock);
+                               read_unlock(&qdisc_tree_lock);
                                goto done;
                        }
                        q_idx++;
                }
-               read_unlock_bh(&qdisc_tree_lock);
+               read_unlock(&qdisc_tree_lock);
        }
 
 done:
@@ -1074,7 +1074,7 @@ static int tc_dump_tclass(struct sk_buff *skb, struct netlink_callback *cb)
        s_t = cb->args[0];
        t = 0;
 
-       read_lock_bh(&qdisc_tree_lock);
+       read_lock(&qdisc_tree_lock);
        list_for_each_entry(q, &dev->qdisc_list, list) {
                if (t < s_t || !q->ops->cl_ops ||
                    (tcm->tcm_parent &&
@@ -1096,7 +1096,7 @@ static int tc_dump_tclass(struct sk_buff *skb, struct netlink_callback *cb)
                        break;
                t++;
        }
-       read_unlock_bh(&qdisc_tree_lock);
+       read_unlock(&qdisc_tree_lock);
 
        cb->args[0] = t;
 
index 6f91518997951cbd13056373f36a98c0f87136b2..88c6a99ce53cfb675c9fb1198ca7f5328be3577e 100644 (file)
    The idea is the following:
    - enqueue, dequeue are serialized via top level device
      spinlock dev->queue_lock.
-   - tree walking is protected by read_lock_bh(qdisc_tree_lock)
+   - tree walking is protected by read_lock(qdisc_tree_lock)
      and this lock is used only in process context.
-   - updates to tree are made under rtnl semaphore or
-     from softirq context (__qdisc_destroy rcu-callback)
-     hence this lock needs local bh disabling.
+   - updates to tree are made only under rtnl semaphore,
+     hence this lock may be made without local bh disabling.
 
    qdisc_tree_lock must be grabbed BEFORE dev->queue_lock!
  */
@@ -57,14 +56,14 @@ DEFINE_RWLOCK(qdisc_tree_lock);
 
 void qdisc_lock_tree(struct net_device *dev)
 {
-       write_lock_bh(&qdisc_tree_lock);
+       write_lock(&qdisc_tree_lock);
        spin_lock_bh(&dev->queue_lock);
 }
 
 void qdisc_unlock_tree(struct net_device *dev)
 {
        spin_unlock_bh(&dev->queue_lock);
-       write_unlock_bh(&qdisc_tree_lock);
+       write_unlock(&qdisc_tree_lock);
 }
 
 /* 
@@ -483,20 +482,6 @@ void qdisc_reset(struct Qdisc *qdisc)
 static void __qdisc_destroy(struct rcu_head *head)
 {
        struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
-       struct Qdisc_ops  *ops = qdisc->ops;
-
-#ifdef CONFIG_NET_ESTIMATOR
-       gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
-#endif
-       write_lock(&qdisc_tree_lock);
-       if (ops->reset)
-               ops->reset(qdisc);
-       if (ops->destroy)
-               ops->destroy(qdisc);
-       write_unlock(&qdisc_tree_lock);
-       module_put(ops->owner);
-
-       dev_put(qdisc->dev);
        kfree((char *) qdisc - qdisc->padded);
 }
 
@@ -504,32 +489,23 @@ static void __qdisc_destroy(struct rcu_head *head)
 
 void qdisc_destroy(struct Qdisc *qdisc)
 {
-       struct list_head cql = LIST_HEAD_INIT(cql);
-       struct Qdisc *cq, *q, *n;
+       struct Qdisc_ops  *ops = qdisc->ops;
 
        if (qdisc->flags & TCQ_F_BUILTIN ||
-               !atomic_dec_and_test(&qdisc->refcnt))
+           !atomic_dec_and_test(&qdisc->refcnt))
                return;
 
-       if (!list_empty(&qdisc->list)) {
-               if (qdisc->ops->cl_ops == NULL)
-                       list_del(&qdisc->list);
-               else
-                       list_move(&qdisc->list, &cql);
-       }
-
-       /* unlink inner qdiscs from dev->qdisc_list immediately */
-       list_for_each_entry(cq, &cql, list)
-               list_for_each_entry_safe(q, n, &qdisc->dev->qdisc_list, list)
-                       if (TC_H_MAJ(q->parent) == TC_H_MAJ(cq->handle)) {
-                               if (q->ops->cl_ops == NULL)
-                                       list_del_init(&q->list);
-                               else
-                                       list_move_tail(&q->list, &cql);
-                       }
-       list_for_each_entry_safe(cq, n, &cql, list)
-               list_del_init(&cq->list);
+       list_del(&qdisc->list);
+#ifdef CONFIG_NET_ESTIMATOR
+       gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
+#endif
+       if (ops->reset)
+               ops->reset(qdisc);
+       if (ops->destroy)
+               ops->destroy(qdisc);
 
+       module_put(ops->owner);
+       dev_put(qdisc->dev);
        call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 }
 
@@ -549,15 +525,15 @@ void dev_activate(struct net_device *dev)
                                printk(KERN_INFO "%s: activation failed\n", dev->name);
                                return;
                        }
-                       write_lock_bh(&qdisc_tree_lock);
+                       write_lock(&qdisc_tree_lock);
                        list_add_tail(&qdisc->list, &dev->qdisc_list);
-                       write_unlock_bh(&qdisc_tree_lock);
+                       write_unlock(&qdisc_tree_lock);
                } else {
                        qdisc =  &noqueue_qdisc;
                }
-               write_lock_bh(&qdisc_tree_lock);
+               write_lock(&qdisc_tree_lock);
                dev->qdisc_sleeping = qdisc;
-               write_unlock_bh(&qdisc_tree_lock);
+               write_unlock(&qdisc_tree_lock);
        }
 
        if (!netif_carrier_ok(dev))