Re: [Xen-devel] xen kernel: build failure

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 28.11.11 at 13:16, Jean Guyader <jean.guyader@xxxxxxxxxxxxx> wrote:
> On 28/11 11:59, Jean Guyader wrote:
>> On 23/11 04:41, Jan Beulich wrote:
>> > >>> On 23.11.11 at 17:31, Jean Guyader <jean.guyader@xxxxxxxxx> wrote:
>> > > On 25 October 2011 13:35, Keir Fraser <keir.xen@xxxxxxxxx> wrote:
>> > >> On 25/10/2011 12:08, "Christoph Egger" <Christoph.Egger@xxxxxxx> wrote:
>> > >>
>> > >>>
>> > >>> Hi,
>> > >>>
>> > >>> Compiling the xen kernel fails with:
>> > >>>
>> > >>> xen/arch/x86/domain.c: In function 'alloc_domain_struct'
>> > >>> xen/arch/x86/domain.c:191: error: negative width in bit-field 
>> > >>> '<anonymous>'
>> > >>
>> > >> Problem is that struct domain has grown bigger than a page for some 
> reason,
>> > >> in your build environment.
>> > >>
>> > >> I can't reproduce this.
>> > >>
>> > >>> Removing the line
>> > >>>
>> > >>> BUILD_BUG_ON(sizeof(*d) > PAGE_SIZE);
>> > >>>
>> > >>> makes xen kernel compile again.
>> > >>
>> > >> But not actually work properly. We only allocate a single page for the
>> > >> domain struct. If the struct is bigger than a page, you'll get memory
>> > >> corruption at run time.
>> > >>
>> > > 
>> > > Is there a reason for that?
>> > 
>> > Of course there is: Post-boot there shouldn't be any allocations of
>> > order greater than zero. This is a generally advisable thing, given that
>> > Xen can't reclaim memory arbitrarily from domains, and in particular
>> > also necessary for tmem.
>> > 
>> > > What would be the recommended to add something
>> > > into the struct domain now if we can't make it bigger than a page.
>> > 
>> > Put a pointer to your data (or larger parts that are already there)
>> > instead of the data itself into struct domain, and allocate it
>> > separately. (You may have seen a patch from Olaf Hering late
>> > yesterday or earlier today that moved the mem_event pieces out
>> > of there for this very reason.)
>> > 
>> > 
>> 
>> I've printed sizeof (struct domain) on boot with xen-unstable and it's
>> exactly a PAGE big. That mean that if I want a add a value to the struct
>> domain (event a pointer) I will have to make some room first.
>> 
> 
> I can see two candidates, that could be turned into pointers in arch_domain
> for x86:
> 
>     /* nestedhvm: translate l2 guest physical to host physical */
>     struct p2m_domain *nested_p2m[MAX_NESTEDP2M];
>     mm_lock_t nested_p2m_lock;
> or
>     struct PITState vpit;
> 
> Let me know which one I should pick. Maybe I should modify both to save
> some space for later.

As written earlier, I think you should leverage Olaf Hering's patch to
split out the three "struct mem_event_domain" instances. Other
candidates would be the watchdog related items or the evtchn array
(the latter would need to be converted to a simple pointer anyway
at some point in order to support more than 4096 event channels
per domain).

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel