Re: [BUG] Inconsistent lock state in virtnet poll

[Thomas Gleixner <tglx@linutronix.de>]

"Michael S. Tsirkin" <mst@redhat.com> writes:
> On Tue, May 05, 2020 at 02:08:56PM +0200, Thomas Gleixner wrote:
>> 
>> The following lockdep splat happens reproducibly on 5.7-rc4
>
>> ================================
>> WARNING: inconsistent lock state
>> 5.7.0-rc4+ #79 Not tainted
>> --------------------------------
>> inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} usage.
>> ip/356 [HC0[0]:SC1[1]:HE1:SE0] takes:
>> f3ee4cd8 (&syncp->seq#2){+.?.}-{0:0}, at: net_rx_action+0xfb/0x390
>> {SOFTIRQ-ON-W} state was registered at:
>>   lock_acquire+0x82/0x300
>>   try_fill_recv+0x39f/0x590
>
> Weird. Where does try_fill_recv acquire any locks?

  u64_stats_update_begin(&rq->stats.syncp);

That's a 32bit kernel which uses a seqcount for this. sequence counts
are "lock" constructs where you need to make sure that writers are
serialized.

Actually the problem at hand is that try_fill_recv() is called from
fully preemptible context initialy and then from softirq context.

Obviously that's for the open() path a non issue, but lockdep does not
know about that. OTOH, there is other code which calls that from
non-softirq context.

The hack below made it shut up. It's obvioulsy not ideal, but at least
it let me look at the actual problem I was chasing down :)

Thanks,

        tglx

8<-----------
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -1243,9 +1243,11 @@ static bool try_fill_recv(struct virtnet
 			break;
 	} while (rq->vq->num_free);
 	if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
+		local_bh_disable();
 		u64_stats_update_begin(&rq->stats.syncp);
 		rq->stats.kicks++;
 		u64_stats_update_end(&rq->stats.syncp);
+		local_bh_enable();
 	}
 
 	return !oom;