Re: [Aoetools-discuss] [BULK] Re: [BULK] aoe6-41 fixes offset trouble (was: Trouble with vblade)

[Sam H. <sa...@co...>]

>> Whoops - I meant to notify the list about this, but forgot.  Turns out
>> it wasn't related to Xen at all, but was a bug that was tripped when
>> using network cards that don't support "scatter-gather" I/O.  It's
>> surprising we got this far without noticing it, but this feature is supported
>> in all but a few 10/100 nics and ... who would be doing AoE over 100Mbit
>> connection anyway?!?  :)
>
> I faced a problem this morning when I upgraded my kernel from 2.6.18 to 
>   2.6.19.1. So I tried to upgrade to aoe6-41. With this driver I was 
> able to partition an AOE drive and to format partitions.
> 
> Unfortunately when I tried to make a copy of my local partition:
> 
> xfsdump -l0 - /dev/sda1 | xfsrestore -p60 - ./
> 
> I got:
> kernel: Bad page state in process 'xfsrestore'
> 
> and everything frees and I got a kernel dump.

That's the one we're talking about now.  Without scatter-gather the
skb layer will increment, then decrement the reference count on pages
in the frag array as part of linearizing the skb.  XFS is handing us
bio pages with a zero reference count -- kmalloc'd pages, we're told
-- which causes the page to be released when the ref count goes back
to zero.

At this point it looks like a bug in XFS as it's the only one in
possession of the knowledge that it's faking out pages in its bios.
Some Hellwigs don't agree.  We're still discussing it.  Right now
you'll have to either stick to aoe6-22 (which make extra copies on
write instead of using the skb frag array) or use a network card that
supports scatter-gather (ethtool -k ethX).

> P.S. Is it possible to setup a Mailing list in such a way that when you 
> press a "reply" button you send a message to the list and not to the 
> author of a mail?

I'll look at sourceforge to see what options we have, but that seems
like it should be automatic.  I've manually modified the headers
in this message for you. :)

Cheers,

Sam