Enabling IOAPIC kills DHCP

[lerdmann]

System: Dual core Intel CPU

Have tried with 2 different guests:
1) eComStation with Warp4 kernel, 1 core
2) eComStation with SMP kernel, 2 cores

In both cases I had enabled NAT so that I would have Internet Access from the guest and therefore the guest needs to use DHCP. If I activate IOAPIC the DHCP Client will not get served an IP address. If I deactivate IOAPIC the DHCP Client will get served an IP address and Internet Access will work.
Obviously, deactivating IOAPIC will not allow me to use 2 cores for 2).

I am attaching the log for 1).

[lerdmann]

Warp 4 kernel and using NAT (DHCP enabled)

[lerdmann]

Potentially related to #12369 ?

[Frank Mehnert]

The IOAPIC is very unlikely the cause for the different behavior of your VMs, at least I cannot imagine this a VBox bug. The log clearly shows that the guest gets 10.0.2.15 assigned...

[lerdmann]

1) Have you looked at #12369 ? Judging from his log he has the same problem that I have. Even though the VirtualBox Version he uses is different.

2) It's true, the log says that the DHCP Server served an address. But maybe the response does not reach the Client ?

3) It's a simple as I stated: turn off IOAPIC-> DHCP Client receives an address, turn on IOAPIC-> DHCP Client does not receive an address. This is always true and don't change anything else but that. Maybe the IOAPIC handling code has some unwanted side effects ? I only have OS/2 guest, therefore I don't know if the behaviour is different for other guests.

Is there any way to ping the DHCP Server from within the guest ? What IP address does it have ? Maybe it is necessary to set up some static route ?

[Valery Ushakov]

Please, can you provide packet capture with --nictrace1=on (see this example). Ideally, captures for both working and failing scenarios.

[lerdmann]

pcap file with IOAPIC enabled

[lerdmann]

pcap file without IOAPIC enabled

[lerdmann]

I have now attached the files.

[Valery Ushakov]

In your bad.pcap file I can see DHCP traffic going into both directions. Apparently the guest doesn't see any of the packets sent to it, so it tries to resend its request, then falls back to do full discovery.

I know nothing about OS/2 and its supported hardware, etc, but I'd hazard a guess - may be the driver doesn't handle RX interrupts correctly with IOAPIC? Have you tried selecting another emulated network adapter? From your log I can see that you use e1000. When a new OS/2 VM is created VirtualBox uses PCnet-FAST III (Am79C973) by default. I don't want to read too much into that selection, but I assume it was done for a reason.

[bird]

Interesting issue, though it could be down to OS/2 configuration (os2apic.psd). Well take a look at it before long.

In the mean time, would be interesting to know whether the intel card works better.

[michaln]

I think what the previous comment meant was knowing whether the AMD PCnet NIC emulation works better... does it?

Also, what exactly is "eComStation" in this context? As far as I know there's more than one version. At any rate, I don't know how interested we're in fixing this if it turns out that the problem can't be reproduced with IBM's ACP2 (does anyone know?).

[michaln]

Oh, and could you please explain why you think it's a good idea to choose the ICH9 chipset type for OS/2 guests? Is the problem reproducible when not using ICH9?

[michaln]

For the record, I verified that the E1000 emulation works fine with an ACP2 guest running the 14.089_SMP kernel and using 2 VCPUs (I/O APIC enabled, which is required for this configuration). This is in a VM which is otherwise using defaults for the OS/2 4.5 guest type.

[abwillis]

Just to confirm, you are able to use 2 CPUs? If I enable IOAPIC in my OS/2 guest it will use much CPU (with a heavy IRQ load) and eventually reboot itself (maybe minutes, hours, or even a day but normally at most a couple of hours), whereupon it may reboot to the same condition or it may hang right after the desktop starts to populate or it may end up giving an internal processing error. This is regardless of whether os2apic.psd or acpi.psd is used. Another quirk with IOAPIC enabled is that the SATA controller may work or it may say the HDD could not be operated, it is about 50/50 whereas it works fine if IOAPIC is not selected. Host: RHEL 6.5 os2krnl: 14.106_SMP One very noticeable thing about its being slow, the clock quickly loses time... within a short time it is a couple minutes behind. As I recall, there was a part of the os2additions that I installed that fixed the slow clock issue when ioapic is not enabled but it isn't helping with ioapic enabled.

[michaln]

Please stick to the topic. If you have a different problem, create a different ticket and describe the problem and the exact configuration in detail. That goes especially for guest configurations that clearly didn't come from IBM. (Or to put it differently: If I have to guess, I'm not going to.)

[michaln]

Given the lack of response, I'll close this as a duplicate of #7736 (problem caused by incomplete ICH9 chipset emulation, does not happen with default VM settings).