KVM-74 and network timeout?

KVM-74 and network timeout?

[Sean Mahrt]

So I’ve got a Gentoo 2.6.25-gentoo-r6 host with virtio.

I’ve got a few guests (all gentoo) with a mounted nfs partition (mythtv).
I’ve got some serious NFS traffic as the guests are doing commercial
detection and transcoding.  I’ve noticed the guest with a lot of Disk I/O
(commercial detection) after a while has a lot of NFS timeouts….  Virtio
or E1000 give me the same result.  Now the real bad part, I’m getting
pings in the order of ms, like 20-100ms on a bridged connection… and NFS
is going crazy...

I’m using libvirt (or sortof) and can post whatever is needed for command.
 I’m using smp on the guests (and the host), and 2.6.25 on the guests…

Where should I start looking?  Is this a KVM-74 issue?  Bump to KVM-75?


Sean

Re: KVM-74 and network timeout?

[Felix Leimbach]

Sean Mahrt wrote:
> I’ve noticed the guest with a lot of Disk I/O
> (commercial detection) after a while has a lot of NFS timeouts….  Virtio
> or E1000 give me the same result.

I noticed exactly the same problem after moving from kvm-64 on a
2.6.25.3 host to kvm-74 on a 2.6.26.3 host.

Adding to your observations:
- CIFS shares are affected as well: Under heavy traffic I get timeouts
from the server, see [1]
- ne2k_pci and rtl8139 guests seem to be affected as well

> Now the real bad part, I’m getting
> pings in the order of ms, like 20-100ms on a bridged connection… and NFS
> is going crazy...

My pings also increased from < 0.1ms to 16ms when the physical interface
of the bridge was maxed out.
Don't know, whether transferring from VM to VM would also trigger that.

>  I’m using smp on the guests (and the host), and 2.6.25 on the guests…

My guests where UP and mostly Windows Server 2003 and one Gentoo 2.6.26,
so I think the culprit is elsewhere.

> Where should I start looking?  Is this a KVM-74 issue?  Bump to KVM-75?

You might try kvm-64 which is rock-solid for me when paired with a
2.6.25 KVM kernel module.

[1]
Sep 6 17:42:58 kernsrc@obelix CIFS VFS: server not responding
Sep 6 17:42:58 kernsrc@obelix CIFS VFS: No response to cmd 46 mid 30836
Sep 6 17:42:58 kernsrc@obelix CIFS VFS: Send error in read = -11
Sep 6 17:51:28 kernsrc@obelix CIFS VFS: server not responding
Sep 6 17:51:28 kernsrc@obelix CIFS VFS: No response for cmd 50 mid 30850

Re: KVM-74 and network timeout?

[Daniel Godás]

[-- Attachment #1: Type: text/plain, Size: 989 bytes --]

Hi, ive been using a JetFlash usb-storage device with kvm-qemu and
Windows XP as a guest. I have seen that before setting up the device
properly, Windows resets it 5 or 6 times, which makes the process
awfully slow. Diving into the code I found that qemu's emulation was
not giving the host the right status when there was a "babble" or
"stall" situation. After applying the following patch the setup time
was cut up by a ~200%.

Cheers,
Daniel

--- old/kvm-75/qemu/usb-linux.c	2008-09-07 18:38:33.000000000 +0200
+++ new/kvm-75/qemu/usb-linux.c	2008-09-18 15:13:47.000000000 +0200
@@ -208,9 +208,13 @@
                 p->len = aurb->urb.actual_length;
                 break;

+            case -EOVERFLOW:
+	    	p->len = USB_RET_BABBLE;
+		break;
             case -EPIPE:
                 set_halt(s, p->devep);
-                /* fall through */
+                p->len = USB_RET_STALL;
+		break;
             default:
                 p->len = USB_RET_NAK;
                 break;

[-- Attachment #2: usb_fix.patch --]
[-- Type: application/octet-stream, Size: 544 bytes --]

--- old/kvm-75/qemu/usb-linux.c	2008-09-07 18:38:33.000000000 +0200
+++ new/kvm-75/qemu/usb-linux.c	2008-09-18 15:13:47.000000000 +0200
@@ -208,9 +208,13 @@
                 p->len = aurb->urb.actual_length;
                 break;
 
+            case -EOVERFLOW:
+	    	p->len = USB_RET_BABBLE;
+		break;
             case -EPIPE:
                 set_halt(s, p->devep);
-                /* fall through */
+                p->len = USB_RET_STALL;
+		break;
             default:
                 p->len = USB_RET_NAK;
                 break;