Re: iproute2: tc deletion freezes whole server

[Cong Wang <xiyou.wangcong@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 1097 bytes --]

On Wed, Mar 25, 2020 at 4:28 AM Václav Zindulka
<vaclav.zindulka@tlapnet.cz> wrote:
>
> On Tue, Mar 24, 2020 at 11:57 PM Cong Wang <xiyou.wangcong@gmail.com> wrote:
> > Hm, my bad, please also run `perf report -g` after you record them,
> > we need the text output with stack traces.
>
> No problem. I've created reports on two servers with different cards.
> See here https://github.com/zvalcav/tc-kernel/tree/master/20200325

That is great!

Your kernel log does not show anything useful, so it did not lead to
any kernel hang or crash etc. at all. (This also means you do not need
to try kdump.)

Are you able to test an experimental patch attached in this email?

It looks like your kernel spent too much time in fq_codel_reset(),
most of it are unnecessary as it is going to be destroyed right after
resetting.

Note: please do not judge the patch, it is merely for testing purpose.
It is obviously ugly and is only a proof of concept. A complete one
should be passing a boolean parameter down to each ->reset(),
but it would be much larger.

Thanks for testing!

[-- Attachment #2: fq_codel_reset.diff --]
[-- Type: application/octet-stream, Size: 585 bytes --]

diff --git a/net/sched/sch_fq_codel.c b/net/sched/sch_fq_codel.c
index 968519ff36e9..df781fd4ff56 100644
--- a/net/sched/sch_fq_codel.c
+++ b/net/sched/sch_fq_codel.c
@@ -337,6 +337,15 @@ static void fq_codel_reset(struct Qdisc *sch)
 	struct fq_codel_sched_data *q = qdisc_priv(sch);
 	int i;
 
+	if (refcount_read(&sch->refcnt) == 0) {
+		for (i = 0; i < q->flows_cnt; i++) {
+			struct fq_codel_flow *flow = q->flows + i;
+
+			fq_codel_flow_purge(flow);
+		}
+		return;
+	}
+
 	INIT_LIST_HEAD(&q->new_flows);
 	INIT_LIST_HEAD(&q->old_flows);
 	for (i = 0; i < q->flows_cnt; i++) {