net_sched: prio: insure proper transactional behavior
authorEric Dumazet <edumazet@google.com>
Mon, 13 Jun 2016 18:33:32 +0000 (11:33 -0700)
committerDavid S. Miller <davem@davemloft.net>
Wed, 15 Jun 2016 19:29:54 +0000 (12:29 -0700)
Now prio_init() can return -ENOMEM, it also has to make sure
any allocated qdiscs are freed, since the caller (qdisc_create()) wont
call ->destroy() handler for us.

More generally, we want a transactional behavior for "tc qdisc
change ...", so prio_tune() should not make modifications if
any error is returned.

It means that we must validate parameters and allocate missing qdisc(s)
before taking root qdisc lock exactly once, to not leave the prio qdisc
in an intermediate state.

Fixes: cbdf45116478 ("net_sched: prio: properly report out of memory errors")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Cong Wang <xiyou.wangcong@gmail.com>
Acked-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/sched/sch_prio.c

index 071718bccdab65da685eae239816ac07715b139d..a356450b747ba57552d6c1e1e35255935580f02a 100644 (file)
@@ -172,8 +172,9 @@ prio_destroy(struct Qdisc *sch)
 static int prio_tune(struct Qdisc *sch, struct nlattr *opt)
 {
        struct prio_sched_data *q = qdisc_priv(sch);
+       struct Qdisc *queues[TCQ_PRIO_BANDS];
+       int oldbands = q->bands, i;
        struct tc_prio_qopt *qopt;
-       int i;
 
        if (nla_len(opt) < sizeof(*qopt))
                return -EINVAL;
@@ -187,54 +188,42 @@ static int prio_tune(struct Qdisc *sch, struct nlattr *opt)
                        return -EINVAL;
        }
 
+       /* Before commit, make sure we can allocate all new qdiscs */
+       for (i = oldbands; i < qopt->bands; i++) {
+               queues[i] = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
+                                             TC_H_MAKE(sch->handle, i + 1));
+               if (!queues[i]) {
+                       while (i > oldbands)
+                               qdisc_destroy(queues[--i]);
+                       return -ENOMEM;
+               }
+       }
+
        sch_tree_lock(sch);
        q->bands = qopt->bands;
        memcpy(q->prio2band, qopt->priomap, TC_PRIO_MAX+1);
 
-       for (i = q->bands; i < TCQ_PRIO_BANDS; i++) {
+       for (i = q->bands; i < oldbands; i++) {
                struct Qdisc *child = q->queues[i];
-               q->queues[i] = &noop_qdisc;
-               if (child != &noop_qdisc) {
-                       qdisc_tree_reduce_backlog(child, child->q.qlen, child->qstats.backlog);
-                       qdisc_destroy(child);
-               }
-       }
-       sch_tree_unlock(sch);
 
-       for (i = 0; i < q->bands; i++) {
-               struct Qdisc *child;
+               qdisc_tree_reduce_backlog(child, child->q.qlen,
+                                         child->qstats.backlog);
+               qdisc_destroy(child);
+       }
 
-               if (q->queues[i] != &noop_qdisc)
-                       continue;
+       for (i = oldbands; i < q->bands; i++)
+               q->queues[i] = queues[i];
 
-               child = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
-                                         TC_H_MAKE(sch->handle, i + 1));
-               if (!child)
-                       return -ENOMEM;
-               sch_tree_lock(sch);
-               q->queues[i] = child;
-               sch_tree_unlock(sch);
-       }
+       sch_tree_unlock(sch);
        return 0;
 }
 
 static int prio_init(struct Qdisc *sch, struct nlattr *opt)
 {
-       struct prio_sched_data *q = qdisc_priv(sch);
-       int i;
-
-       for (i = 0; i < TCQ_PRIO_BANDS; i++)
-               q->queues[i] = &noop_qdisc;
-
-       if (opt == NULL) {
+       if (!opt)
                return -EINVAL;
-       } else {
-               int err;
 
-               if ((err = prio_tune(sch, opt)) != 0)
-                       return err;
-       }
-       return 0;
+       return prio_tune(sch, opt);
 }
 
 static int prio_dump(struct Qdisc *sch, struct sk_buff *skb)