Re: KVM induced panic on 2.6.38[2367] & 2.6.39

Le lundi 22 août 2011 à 09:36 +0300, Avi Kivity a écrit :

On 08/20/2011 04:16 PM, Brad Campbell wrote:

quoted

Author: Alexander Duyck [off-list ref]
Date:   Thu Jul 1 13:28:27 2010 +0000

    x86: Drop CONFIG_MCORE2 check around setting of NET_IP_ALIGN

    This patch removes the CONFIG_MCORE2 check from around 
NET_IP_ALIGN.  It is
    based on a suggestion from Andi Kleen.  The assumption is that 
there are
    not any x86 cores where unaligned access is really slow, and this 
change
    would allow for a performance improvement to still exist on 
configurations
    that are not necessarily optimized for Core 2.

    Cc: Andi Kleen [off-list ref]
    Cc: Thomas Gleixner [off-list ref]
    Cc: Ingo Molnar [off-list ref]
    Cc: "H. Peter Anvin" [off-list ref]
    Cc: x86@kernel.org
    Signed-off-by: Alexander Duyck [off-list ref]
    Signed-off-by: Jeff Kirsher [off-list ref]
    Acked-by: H. Peter Anvin [off-list ref]
    Signed-off-by: David S. Miller [off-list ref]

:040000 040000 5a15867789080a2f67a74b17c4422f85b7a9fb4a 
b98769348bd765731ca3ff03b33764257e23226c M    arch

I can confirm this bug exists in the 3.0 kernel, however I'm unable to 
reproduce it on todays git.

So anyone using netfilter, kvm and bridge on kernels between 
2.6.36-rc1 and 3.0 may hit this bug, but it looks like it is fixed in 
the current 3.1-rc kernels.

Thanks for this effort.  I don't think this patch is buggy in itself, it 
merely exposed another bug which was fixed later on.

Some piece of hardware has a 2-byte offset requirement, and driver
incorrectly assumed NET_IP_ALIGN was 2 on x86.

Brad, could you post your config (lsmod, dmesg) again ?

tg3.c code for example uses a private value, not related to NET_IP_ALIGN

#define TG3_RAW_IP_ALIGN 2