Re: [kvm-devel] [GIT PULL][RESEND] Late KVM Updates for the 2.6.23 merge window

[Avi Kivity <avi@qumranet.com>]

Linus Torvalds wrote:
> On Sat, 21 Jul 2007, S.Çağlar Onur wrote:
>   
>> With Linus's latest git, shutting down a guest (fired with -smp 2 -m 512) sometimes 
>> ends up like [1], this occured as soon as qemu window closed.
>>
>> [1] http://cekirdek.pardus.org.tr/~caglar/kvm/dmesg.latest
>>     
>
> [  737.460654] Bad page state in process 'qemu-kvm'
> [  737.460656] page:f5e68000 flags:0xea020000 mapping:00000000 mapcount:2 count:0
> [  737.460657] Trying to fix it up, but a reboot is needed
> [  737.460659] Backtrace:
> [  737.460691]  [<c0159e70>] bad_page+0x64/0x8e
> [  737.460733]  [<c015a7ee>] free_hot_cold_page+0x68/0x15a
>
> That's the "free_pages_check()", and in particular it seems to be 
> "page_mapcount()" being non-zero that triggered that thing.
>
> So it looks like something in KVM isn't coherent about the mapping vs the 
> usage counters..
>
>   

Far more prosaic.  It calls __free_page() on the cleanup path with the
page's virtual address instead of the struct page.  I guess
__free_page() is happy with a zeroed out blob, which is the most likely
contents.

Trivial patch shortly.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.