Virtualization - Feb 2011 - [PATCH] virtio-net: add schedule check to napi_enable call in refill_work

Justification:

Impact: Under heavy network I/O load virtio-net driver crashes making VM guest
unusable.

Testcases:

1) Sergey Svishchev reports that servers that run java webapps with high java
heap usage (especially when heap size is close to physical memory size) helps
trigger one of aforementioned bugs. Unfortunately, I don't have a simple
test case.

2) Peter Lieven reports that his binary NNTP newsfeed test servers crash without
this patch.

3) Bruce Rogers of Novell has asked for this patch to be integrated but the
request was mysteriously ignored.  Is is purported that this patch is being
distributed with SLES.

4) I can crash 2.6.32 and 2.6.38-rc4 by simply running "scp -r
/nfs/read-only/1 otherhost:/target/1" and "scp -r /nfs/read-only/2
otherhost:/target/2" concurrently with a mix of small to medium files for a
few hours usually.  I've never seen more than 200 GB copied before the crash
occurs.  Both 2.6.32 and 2.6.38-rc4 with this patch will copy more than 200GB
unfailingly this way.

See
https://bugs.launchpad.net/bugs/579276
for more details.



--- drivers/net/virtio_net.c.orig	2011-02-08 14:34:51.444099190 -0500
+++ drivers/net/virtio_net.c	2011-02-08 14:18:00.484400134 -0500
@@ -446,6 +446,20 @@
 	}
 }
 
+static void virtnet_napi_enable(struct virtnet_info *vi)
+{
+	napi_enable(&vi->napi);
+
+	/* If all buffers were filled by other side before we napi_enabled, we
+	 * won't get another interrupt, so process any outstanding packets
+	 * now.  virtnet_poll wants re-enable the queue, so we disable here.
+	 * We synchronize against interrupts via NAPI_STATE_SCHED */
+	if (napi_schedule_prep(&vi->napi)) {
+		virtqueue_disable_cb(vi->rvq);
+		__napi_schedule(&vi->napi);
+	}
+}
+
 static void refill_work(struct work_struct *work)
 {
 	struct virtnet_info *vi;
@@ -454,7 +468,7 @@
 	vi = container_of(work, struct virtnet_info, refill.work);
 	napi_disable(&vi->napi);
 	still_empty = !try_fill_recv(vi, GFP_KERNEL);
-	napi_enable(&vi->napi);
+	virtnet_napi_enable(vi);
 
 	/* In theory, this can happen: if we don't get any buffers in
 	 * we will *never* try to fill again. */
@@ -638,16 +652,7 @@
 {
 	struct virtnet_info *vi = netdev_priv(dev);
 
-	napi_enable(&vi->napi);
-
-	/* If all buffers were filled by other side before we napi_enabled, we
-	 * won't get another interrupt, so process any outstanding packets
-	 * now.  virtnet_poll wants re-enable the queue, so we disable here.
-	 * We synchronize against interrupts via NAPI_STATE_SCHED */
-	if (napi_schedule_prep(&vi->napi)) {
-		virtqueue_disable_cb(vi->rvq);
-		__napi_schedule(&vi->napi);
-	}
+	virtnet_napi_enable(vi);
 	return 0;
 }

On Thu, 10 Feb 2011 06:59:25 am Ken Stailey wrote:
> Justification:
> 
> Impact: Under heavy network I/O load virtio-net driver crashes making VM
guest unusable.
Hmm, this went badly wrong.  I acked this patch, and it was mailed to
netdev six months ago.

Bruce's patch used spaces instead of tabs, but that should not have caused
it to be dropped.  I've taken that and ported it forwards, will repost now.

Thanks for picking this up off the floor!
Rusty.