HVM domain crash when network attach

HVM domain crash when network attach

[Hirofumi Tsujimura]

Hi all,

I found a problem when I tested vnif of PV-on-HVM.
My operation for the test is following.

On domain0:
# xm create -f domainHVM.config
# xm network-attach domainHVM

Then, the kernel of domainHVM panicked and outputted a message
"kernel panic - not syncing: Unable to reduce memory reservation".

The kernel sometimes panicked not always when I tested it some time.
In my investigation

### netfront/netfront.c
285 err = xenbus_scanf(XBT_NIL, dev->otherend, "feature-rx-copy", "%u",
286                    &feature_rx_copy);  <---------------- (A)
287 if (err != 1)
288   feature_rx_copy = 0;
===
802 if ( nr_flips != 0 ) {  <------------------------------- (B)
===
827   /* Check return status of HYPERVISOR_memory_op(). */
828   if (unlikely(np->rx_mcl[i].result != i))
829     panic("Unable to reduce memory reservation\n"); <------ (C)
830   } else {
831     if (HYPERVISOR_memory_op(XENMEM_decrease_reservation,
832                                                 &reservation) != i)
833       panic("Unable to reduce memory reservation\n"); <------- (D)
834 }

### netback/xenbus.c
111 err = xenbus_printf(xbt, dev->nodename,
                                  "feature-rx-copy", "%d", 1);<-- (E)

The problem happans in (C) or (D) when (B) become true.
I think that (B) must not become true on a domainHVM.
Because a domainHVM only uses rx_copy mode.

(B) is true because (A) fails.
(A) fails because (E) is executed after (A).
(E) must be execute before (A).
However (E) is not always executed before (A).

The test environment is as follows.
  - Architecture  : x86_32 with VTx
  - Xen change set: 11751
  - Domain 0      : kernel2.6.16.29-xen
  - Domain U      : kernel2.6.16.29-xen
  - Domain VTx    : kernel2.6.16

Does anyone know how to fix the problem ?
Thank you.
--
FUJITSU COMPUTER TECHNOLOGIES LIMITED
Development Dept. 5  Division 1
Hirofumi Tsujimura
E-MAIL: tsujimura.hirof@jp.fujitsu.com

Re: HVM domain crash when network attach

[Keir Fraser]

On 13/10/06 14:44, "Hirofumi Tsujimura" <tsujimura.hirof@jp.fujitsu.com>
wrote:

> Then, the kernel of domainHVM panicked and outputted a message
> "kernel panic - not syncing: Unable to reduce memory reservation".

That's a nasty race caused by netfront not waiting for the backend to reach
InitWait state before reading its advertised features. Since this is a domU
guest issue I'd like to have it fixed for 3.0.3. I've put a fix in
xen-unstable as c/s 11755:7da100019e00 and I'll also do a 3.0.3-rc5 shortly,
since this touches some moderately complicated code.

 -- Keir

Re: HVM domain crash when network attach

[Hirofumi Tsujimura]

Hi Keir,

I confirmed that the problem was fixed in the changeset 11760.
Thank you.

-- Tsujimura


Keir Fraser wrote:
> On 13/10/06 14:44, "Hirofumi Tsujimura" <tsujimura.hirof@jp.fujitsu.com>
> wrote:
> 
>> Then, the kernel of domainHVM panicked and outputted a message
>> "kernel panic - not syncing: Unable to reduce memory reservation".
> 
> That's a nasty race caused by netfront not waiting for the backend to reach
> InitWait state before reading its advertised features. Since this is a domU
> guest issue I'd like to have it fixed for 3.0.3. I've put a fix in
> xen-unstable as c/s 11755:7da100019e00 and I'll also do a 3.0.3-rc5 shortly,
> since this touches some moderately complicated code.
> 
>  -- Keir