[PATCH] pkt_sched: fq: do not hold qdisc lock while allocating memory

[PATCH] pkt_sched: fq: do not hold qdisc lock while allocating memory

[Eric Dumazet]

From: Eric Dumazet <edumazet@google.com>

Resizing fq hash table allocates memory while holding qdisc spinlock,
with BH disabled.

This is definitely not good, as allocation might sleep.

We can drop the lock and get it when needed, we hold RTNL so no other
changes can happen at the same time.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Fixes: afe4fd062416 ("pkt_sched: fq: Fair Queue packet scheduler")
---
 net/sched/sch_fq.c |   21 +++++++++++++++------
 1 file changed, 15 insertions(+), 6 deletions(-)

diff --git a/net/sched/sch_fq.c b/net/sched/sch_fq.c
index 08ef7a42c0e4..21e251766eb1 100644
--- a/net/sched/sch_fq.c
+++ b/net/sched/sch_fq.c
@@ -601,6 +601,7 @@ static int fq_resize(struct Qdisc *sch, u32 log)
 {
 	struct fq_sched_data *q = qdisc_priv(sch);
 	struct rb_root *array;
+	void *old_fq_root;
 	u32 idx;
 
 	if (q->fq_root && log == q->fq_trees_log)
@@ -615,13 +616,19 @@ static int fq_resize(struct Qdisc *sch, u32 log)
 	for (idx = 0; idx < (1U << log); idx++)
 		array[idx] = RB_ROOT;
 
-	if (q->fq_root) {
-		fq_rehash(q, q->fq_root, q->fq_trees_log, array, log);
-		fq_free(q->fq_root);
-	}
+	sch_tree_lock(sch);
+
+	old_fq_root = q->fq_root;
+	if (old_fq_root)
+		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
+
 	q->fq_root = array;
 	q->fq_trees_log = log;
 
+	sch_tree_unlock(sch);
+
+	fq_free(old_fq_root);
+
 	return 0;
 }
 
@@ -697,9 +704,11 @@ static int fq_change(struct Qdisc *sch, struct nlattr *opt)
 		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
 	}
 
-	if (!err)
+	if (!err) {
+		sch_tree_unlock(sch);
 		err = fq_resize(sch, fq_log);
-
+		sch_tree_lock(sch);
+	}
 	while (sch->q.qlen > sch->limit) {
 		struct sk_buff *skb = fq_dequeue(sch);
 

Re: [PATCH] pkt_sched: fq: do not hold qdisc lock while allocating memory

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Thu, 06 Mar 2014 22:57:52 -0800

> From: Eric Dumazet <edumazet@google.com>
> 
> Resizing fq hash table allocates memory while holding qdisc spinlock,
> with BH disabled.
> 
> This is definitely not good, as allocation might sleep.
> 
> We can drop the lock and get it when needed, we hold RTNL so no other
> changes can happen at the same time.
> 
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Fixes: afe4fd062416 ("pkt_sched: fq: Fair Queue packet scheduler")

Eric I think you can simplify things a little further, and in fact I
think it makes these functions easier to understand.

The fq_resize() part should just grab the lock around the rehash and
the update of q->fq_root and q->fq_trees_log.

fq_change() should only grab the lock around the fq_dequeue() loop
and the call to qdisc_tree_descrease_qlen().  The rest of this
function is just validating netlink attributes and looking at state
that cannot change while we hold RTNL.

Re: [PATCH] pkt_sched: fq: do not hold qdisc lock while allocating memory

[Eric Dumazet]

On Fri, 2014-03-07 at 17:04 -0500, David Miller wrote:

> Eric I think you can simplify things a little further, and in fact I
> think it makes these functions easier to understand.
> 
> The fq_resize() part should just grab the lock around the rehash and
> the update of q->fq_root and q->fq_trees_log.
> 
> fq_change() should only grab the lock around the fq_dequeue() loop
> and the call to qdisc_tree_descrease_qlen().  The rest of this
> function is just validating netlink attributes and looking at state
> that cannot change while we hold RTNL.

Hmm, but all these parameters we change in fq_change() are read by other
cpus doing enqueue()/dequeue().

They are integers, so a race would be not a big deal I guess, but better
add a fat comment then ;)

Re: [PATCH] pkt_sched: fq: do not hold qdisc lock while allocating memory

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Fri, 07 Mar 2014 15:02:42 -0800

> On Fri, 2014-03-07 at 17:04 -0500, David Miller wrote:
> 
>> Eric I think you can simplify things a little further, and in fact I
>> think it makes these functions easier to understand.
>> 
>> The fq_resize() part should just grab the lock around the rehash and
>> the update of q->fq_root and q->fq_trees_log.
>> 
>> fq_change() should only grab the lock around the fq_dequeue() loop
>> and the call to qdisc_tree_descrease_qlen().  The rest of this
>> function is just validating netlink attributes and looking at state
>> that cannot change while we hold RTNL.
> 
> Hmm, but all these parameters we change in fq_change() are read by other
> cpus doing enqueue()/dequeue().
> 
> They are integers, so a race would be not a big deal I guess, but better
> add a fat comment then ;)

Good point, this patch as-is is fine, so applied.

Re: [PATCH] pkt_sched: fq: do not hold qdisc lock while allocating memory

[Eric Dumazet]

On Sat, 2014-03-08 at 19:10 -0500, David Miller wrote:

> Good point, this patch as-is is fine, so applied.

I see you applied it in net-next, but it should be in net tree

Thanks !

Re: [PATCH] pkt_sched: fq: do not hold qdisc lock while allocating memory

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Sun, 09 Mar 2014 07:18:26 -0700

> On Sat, 2014-03-08 at 19:10 -0500, David Miller wrote:
> 
>> Good point, this patch as-is is fine, so applied.
> 
> I see you applied it in net-next, but it should be in net tree

My bad, I applied it now to 'net' too and queued it up for -stable.